| Message ID | 20260213-kbingham-quantizers-v7-5-1626b9aaabf1@ideasonboard.com |
|---|---|
| State | Accepted |
| Headers | show |
| Series |
|
| Related | show |
Quoting Kieran Bingham (2026-02-14 01:57:44) > Provide tests to validate the conditions of FixedPoint types used > within libcamera explicitly. > > Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com> Reviewed-by: Paul Elder <paul.elder@ideasonboard.com> > > --- > v3: > - Rename quantized_type to QuantizedType > > v5: > > - Use string_view for introduction of type > - move min/max check to static assert > - Squash down all tests into a single implementation and remove extra > type proliferation. > - Use Q/UQ types directly. > - use std::cout consistently > > v6: > - Add notes on test introduction > - Expand Q12.4 to include clamp tests to validate int16_t range usage > - Add storage selection assertion checks > - Move clamp checks to their types to expand > - Add {U,}Q<4,20> to check up to 24 bit precision > > v7: > - Provide signed register extension tests > --- > test/ipa/libipa/fixedpoint.cpp | 209 ++++++++++++++++++++++++++++++++++++++++- > 1 file changed, 205 insertions(+), 4 deletions(-) > > diff --git a/test/ipa/libipa/fixedpoint.cpp b/test/ipa/libipa/fixedpoint.cpp > index 4b017e86a74f..2907426b7a69 100644 > --- a/test/ipa/libipa/fixedpoint.cpp > +++ b/test/ipa/libipa/fixedpoint.cpp > @@ -5,12 +5,14 @@ > * Fixed / Floating point utility tests > */ > > +#include "../src/ipa/libipa/fixedpoint.h" > + > #include <cmath> > #include <iostream> > #include <map> > #include <stdint.h> > > -#include "../src/ipa/libipa/fixedpoint.h" > +#include <libcamera/base/utils.h> > > #include "test.h" > > @@ -95,14 +97,213 @@ protected: > return TestPass; > } > > - int run() > + template<typename Q> > + int quantizedCheck(float input, typename Q::QuantizedType expected, float value) > { > - /* fixed point conversion test */ > - if (testFixedPoint() != TestPass) > + Q q(input); > + using T = typename Q::QuantizedType; > + > + std::cout << " Checking " << input << " == " << q << std::endl; > + > + T quantized = q.quantized(); > + if (quantized != expected) { > + std::cout << " ** Q Expected " << input > + << " to quantize to " << utils::hex(expected) > + << ", got " << utils::hex(quantized) > + << " - (" << q << ")" > + << std::endl; > + return 1; > + } > + > + if ((std::abs(q.value() - value)) > 0.0001f) { > + std::cout << " ** V Expected " << input > + << " to quantize to " << value > + << ", got " << q.value() > + << " - (" << q << ")" > + << std::endl; > + return 1; > + } > + > + return 0; > + } > + > + template<typename Q, typename RegisterType> > + int signExtendCheck() > + { > + Q minValue(Q::TraitsType::min); > + using T = typename Q::QuantizedType; > + T minQuantized = minValue.quantized(); > + > + /* Take the value and expand it into a larger register */ > + RegisterType reg = static_cast<RegisterType>(minValue.quantized()); > + > + std::cout << " Checking " << minValue << " == " << utils::hex(reg) << std::endl; > + > + /* > + * Ensure that the minimum negative value is not sign > + * extended when casting up to a larger register storage > + * type. > + */ > + if (reg != minQuantized) { > + std::cout << " ** Sign extension check failed for min value. Expected " > + << utils::hex(minQuantized) << ", got " > + << utils::hex(reg) << std::endl; > + return 1; > + } > + > + /* And we should be consistent when we convert back to Q */ > + Q q2(static_cast<T>(reg)); > + if (!(q2 == minValue)) { > + std::cout << " ** Sign extension check failed for min value. Expected " > + << minValue << ", got " << q2 << std::endl; > + return 1; > + } > + > + return 0; > + } > + > + template<typename Q> > + void introduce(std::string_view type) > + { > + using T = typename Q::QuantizedType; > + > + std::cout << std::endl; > + > + std::cout << type > + << "(" << Q::TraitsType::min << " .. " << Q::TraitsType::max << ") " > + << " Min: " << Q(Q::TraitsType::min) > + << " -- Max: " << Q(Q::TraitsType::max) > + << " Step:" << Q(T(1)).value() > + << std::endl; > + } > + > + int testFixedPointQuantizers() > + { > + unsigned int fails = 0; > + > + /* > + * These aim to specifically test all the corner cases of the > + * quantization and de-quantization process. Including clamping > + * to min/max, zero points and making sure that steps are > + * correct. > + * > + * In particular test signed and unsigned types and a mix of the > + * highest bit width of a storage type and smaller widths that > + * require bit masking and sign extension. > + * > + * Note we must hard code these. Any calculation of expected > + * values risks replicating bugs in the implementation. > + * > + * As the underlying types are integer and float the limit of > + * precision is around 24 bits so we do not test wider types. > + */ > + > + /* clang-format off */ > + > + /* Q1.7(-1 .. 0.992188) Min: [0x80:-1] -- Max: [0x7f:0.992188] Step:0.0078125*/ > + introduce<Q<1, 7>>("Q1.7"); > + fails += quantizedCheck<Q<1, 7>>(-2.000f, 0b1'0000000, -1.0f); /* Clamped to Min */ > + fails += quantizedCheck<Q<1, 7>>(-1.000f, 0b1'0000000, -1.0f); /* Min */ > + fails += quantizedCheck<Q<1, 7>>(-0.992f, 0b1'0000001, -0.992188f); /* Min + 1 step */ > + fails += quantizedCheck<Q<1, 7>>(-0.006f, 0b1'1111111, -0.0078125f); /* -1 step */ > + fails += quantizedCheck<Q<1, 7>>( 0.000f, 0b0'0000000, 0.0f); /* Zero */ > + fails += quantizedCheck<Q<1, 7>>( 0.008f, 0b0'0000001, 0.0078125f); /* +1 step */ > + fails += quantizedCheck<Q<1, 7>>( 0.992f, 0b0'1111111, 0.992188f); /* Max */ > + fails += quantizedCheck<Q<1, 7>>( 2.000f, 0b0'1111111, 0.992188f); /* Clamped to Max */ > + > + /* UQ1.7(0 .. 1.99219) Min: [0x00:0] -- Max: [0xff:1.99219] Step:0.0078125 */ > + introduce<UQ<1, 7>>("UQ1.7"); > + fails += quantizedCheck<UQ<1, 7>>(-1.0f, 0b0'0000000, 0.0f); /* Clamped to Min */ > + fails += quantizedCheck<UQ<1, 7>>( 0.0f, 0b0'0000000, 0.0f); /* Min / Zero */ > + fails += quantizedCheck<UQ<1, 7>>( 1.0f, 0b1'0000000, 1.0f); /* Mid */ > + fails += quantizedCheck<UQ<1, 7>>( 1.992f, 0b1'1111111, 1.99219f); /* Max */ > + fails += quantizedCheck<UQ<1, 7>>( 2.000f, 0b1'1111111, 1.99219f); /* Clamped to Max */ > + > + /* UQ4.8(0 .. 15.9961) Min: [0x0000:0] -- Max: [0x0fff:15.9961] Step:0.00390625 */ > + introduce<UQ<4, 8>>("UQ4.8"); > + fails += quantizedCheck<UQ<4, 8>>( 0.0f, 0b0000'00000000, 0.00f); > + fails += quantizedCheck<UQ<4, 8>>(16.0f, 0b1111'11111111, 15.9961f); > + > + /* Q5.4(-16 .. 15.9375) Min: [0x0100:-16] -- Max: [0x00ff:15.9375] Step:0.0625 */ > + introduce<Q<5, 4>>("Q5.4"); > + fails += quantizedCheck<Q<5, 4>>(-16.00f, 0b10000'0000, -16.00f); > + fails += quantizedCheck<Q<5, 4>>( 15.94f, 0b01111'1111, 15.9375f); > + > + /* UQ5.8(0 .. 31.9961) Min: [0x0000:0] -- Max: [0x1fff:31.9961] Step:0.00390625 */ > + introduce<UQ<5, 8>>("UQ5.8"); > + fails += quantizedCheck<UQ<5, 8>>( 0.00f, 0b00000'00000000, 0.00f); > + fails += quantizedCheck<UQ<5, 8>>(32.00f, 0b11111'11111111, 31.9961f); > + > + /* Q12.4(-2048 .. 2047.94) Min: [0x8000:-2048] -- Max: [0x7fff:2047.94] Step:0.0625 */ > + introduce<Q<12, 4>>("Q12.4"); > + fails += quantizedCheck<Q<12, 4>>( 0.0f, 0b000000000000'0000, 0.0f); > + fails += quantizedCheck<Q<12, 4>>( 7.5f, 0b000000000111'1000, 7.5f); > + > + /* UQ12.4(0 .. 4095.94) Min: [0x0000:0] -- Max: [0xffff:4095.94] Step:0.0625 */ > + introduce<UQ<12, 4>>("UQ12.4"); > + fails += quantizedCheck<UQ<12, 4>>(0.0f, 0b000000000000'0000, 0.0f); > + fails += quantizedCheck<UQ<12, 4>>(7.5f, 0b000000000111'1000, 7.5f); > + > + /* Q4.20 */ > + introduce<Q<4, 20>>("Q4.20"); > + fails += quantizedCheck<Q<4, 20>>( -9.0f, 0b1000'00000000000000000000, -8.0f); > + fails += quantizedCheck<Q<4, 20>>( -8.0f, 0b1000'00000000000000000000, -8.0f); > + fails += quantizedCheck<Q<4, 20>>( 8.0f, 0b0111'11111111111111111111, 8.0f); > + fails += quantizedCheck<Q<4, 20>>( 9.0f, 0b0111'11111111111111111111, 8.0f); > + > + /* UQ4.20 */ > + introduce<UQ<4, 20>>("UQ4.20"); > + fails += quantizedCheck<UQ<4, 20>>(-1.0f, 0b0000'00000000000000000000, 0.0f); > + fails += quantizedCheck<UQ<4, 20>>( 0.0f, 0b0000'00000000000000000000, 0.0f); > + fails += quantizedCheck<UQ<4, 20>>(16.0f, 0b1111'11111111111111111111, 16.0f); > + fails += quantizedCheck<UQ<4, 20>>(20.0f, 0b1111'11111111111111111111, 16.0f); > + > + /* Storage selection tests */ > + static_assert(std::is_same_v<typename Q<4, 4>::TraitsType::QuantizedType, uint8_t>); > + static_assert(std::is_same_v<typename UQ<4, 4>::TraitsType::QuantizedType, uint8_t>); > + static_assert(std::is_same_v<typename Q<8, 8>::TraitsType::QuantizedType, uint16_t>); > + static_assert(std::is_same_v<typename UQ<8, 8>::TraitsType::QuantizedType, uint16_t>); > + static_assert(std::is_same_v<typename Q<20, 4>::TraitsType::QuantizedType, uint32_t>); > + static_assert(std::is_same_v<typename UQ<20, 4>::TraitsType::QuantizedType, uint32_t>); > + > + /* > + * Test and validate that sign extension can not modify a > + * quantized value when stored or cast to a larger register. > + */ > + std::cout << std::endl; > + std::cout << "Testing sign extension of quantized values when cast to larger registers" << std::endl; > + fails += signExtendCheck<Q<2, 4>, uint8_t>(); > + fails += signExtendCheck<Q<4, 4>, uint8_t>(); > + fails += signExtendCheck<Q<4, 4>, uint16_t>(); > + fails += signExtendCheck<Q<8, 8>, uint16_t>(); > + fails += signExtendCheck<Q<8, 8>, uint32_t>(); > + > + /* clang-format on */ > + > + std::cout << std::endl; > + > + if (fails > 0) { > + cout << "Fixed point quantizer tests failed: " > + << std::dec << fails << " failures." << std::endl; > return TestFail; > + } > > return TestPass; > } > + > + int run() > + { > + unsigned int fails = 0; > + > + /* fixed point conversion test */ > + if (testFixedPoint() != TestPass) > + fails++; > + > + if (testFixedPointQuantizers() != TestPass) > + fails++; > + > + return fails ? TestFail : TestPass; > + } > }; > > TEST_REGISTER(FixedPointUtilsTest) > > -- > 2.52.0 >
Quoting Kieran Bingham (2026-02-13 17:57:44) > Provide tests to validate the conditions of FixedPoint types used > within libcamera explicitly. > > Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com> Looks good to me. Reviewed-by: Stefan Klug <stefan.klug@ideasonboard.com> Regards, Stefan > > --- > v3: > - Rename quantized_type to QuantizedType > > v5: > > - Use string_view for introduction of type > - move min/max check to static assert > - Squash down all tests into a single implementation and remove extra > type proliferation. > - Use Q/UQ types directly. > - use std::cout consistently > > v6: > - Add notes on test introduction > - Expand Q12.4 to include clamp tests to validate int16_t range usage > - Add storage selection assertion checks > - Move clamp checks to their types to expand > - Add {U,}Q<4,20> to check up to 24 bit precision > > v7: > - Provide signed register extension tests > --- > test/ipa/libipa/fixedpoint.cpp | 209 ++++++++++++++++++++++++++++++++++++++++- > 1 file changed, 205 insertions(+), 4 deletions(-) > > diff --git a/test/ipa/libipa/fixedpoint.cpp b/test/ipa/libipa/fixedpoint.cpp > index 4b017e86a74f..2907426b7a69 100644 > --- a/test/ipa/libipa/fixedpoint.cpp > +++ b/test/ipa/libipa/fixedpoint.cpp > @@ -5,12 +5,14 @@ > * Fixed / Floating point utility tests > */ > > +#include "../src/ipa/libipa/fixedpoint.h" > + > #include <cmath> > #include <iostream> > #include <map> > #include <stdint.h> > > -#include "../src/ipa/libipa/fixedpoint.h" > +#include <libcamera/base/utils.h> > > #include "test.h" > > @@ -95,14 +97,213 @@ protected: > return TestPass; > } > > - int run() > + template<typename Q> > + int quantizedCheck(float input, typename Q::QuantizedType expected, float value) > { > - /* fixed point conversion test */ > - if (testFixedPoint() != TestPass) > + Q q(input); > + using T = typename Q::QuantizedType; > + > + std::cout << " Checking " << input << " == " << q << std::endl; > + > + T quantized = q.quantized(); > + if (quantized != expected) { > + std::cout << " ** Q Expected " << input > + << " to quantize to " << utils::hex(expected) > + << ", got " << utils::hex(quantized) > + << " - (" << q << ")" > + << std::endl; > + return 1; > + } > + > + if ((std::abs(q.value() - value)) > 0.0001f) { > + std::cout << " ** V Expected " << input > + << " to quantize to " << value > + << ", got " << q.value() > + << " - (" << q << ")" > + << std::endl; > + return 1; > + } > + > + return 0; > + } > + > + template<typename Q, typename RegisterType> > + int signExtendCheck() > + { > + Q minValue(Q::TraitsType::min); > + using T = typename Q::QuantizedType; > + T minQuantized = minValue.quantized(); > + > + /* Take the value and expand it into a larger register */ > + RegisterType reg = static_cast<RegisterType>(minValue.quantized()); > + > + std::cout << " Checking " << minValue << " == " << utils::hex(reg) << std::endl; > + > + /* > + * Ensure that the minimum negative value is not sign > + * extended when casting up to a larger register storage > + * type. > + */ > + if (reg != minQuantized) { > + std::cout << " ** Sign extension check failed for min value. Expected " > + << utils::hex(minQuantized) << ", got " > + << utils::hex(reg) << std::endl; > + return 1; > + } > + > + /* And we should be consistent when we convert back to Q */ > + Q q2(static_cast<T>(reg)); > + if (!(q2 == minValue)) { > + std::cout << " ** Sign extension check failed for min value. Expected " > + << minValue << ", got " << q2 << std::endl; > + return 1; > + } > + > + return 0; > + } > + > + template<typename Q> > + void introduce(std::string_view type) > + { > + using T = typename Q::QuantizedType; > + > + std::cout << std::endl; > + > + std::cout << type > + << "(" << Q::TraitsType::min << " .. " << Q::TraitsType::max << ") " > + << " Min: " << Q(Q::TraitsType::min) > + << " -- Max: " << Q(Q::TraitsType::max) > + << " Step:" << Q(T(1)).value() > + << std::endl; > + } > + > + int testFixedPointQuantizers() > + { > + unsigned int fails = 0; > + > + /* > + * These aim to specifically test all the corner cases of the > + * quantization and de-quantization process. Including clamping > + * to min/max, zero points and making sure that steps are > + * correct. > + * > + * In particular test signed and unsigned types and a mix of the > + * highest bit width of a storage type and smaller widths that > + * require bit masking and sign extension. > + * > + * Note we must hard code these. Any calculation of expected > + * values risks replicating bugs in the implementation. > + * > + * As the underlying types are integer and float the limit of > + * precision is around 24 bits so we do not test wider types. > + */ > + > + /* clang-format off */ > + > + /* Q1.7(-1 .. 0.992188) Min: [0x80:-1] -- Max: [0x7f:0.992188] Step:0.0078125*/ > + introduce<Q<1, 7>>("Q1.7"); > + fails += quantizedCheck<Q<1, 7>>(-2.000f, 0b1'0000000, -1.0f); /* Clamped to Min */ > + fails += quantizedCheck<Q<1, 7>>(-1.000f, 0b1'0000000, -1.0f); /* Min */ > + fails += quantizedCheck<Q<1, 7>>(-0.992f, 0b1'0000001, -0.992188f); /* Min + 1 step */ > + fails += quantizedCheck<Q<1, 7>>(-0.006f, 0b1'1111111, -0.0078125f); /* -1 step */ > + fails += quantizedCheck<Q<1, 7>>( 0.000f, 0b0'0000000, 0.0f); /* Zero */ > + fails += quantizedCheck<Q<1, 7>>( 0.008f, 0b0'0000001, 0.0078125f); /* +1 step */ > + fails += quantizedCheck<Q<1, 7>>( 0.992f, 0b0'1111111, 0.992188f); /* Max */ > + fails += quantizedCheck<Q<1, 7>>( 2.000f, 0b0'1111111, 0.992188f); /* Clamped to Max */ > + > + /* UQ1.7(0 .. 1.99219) Min: [0x00:0] -- Max: [0xff:1.99219] Step:0.0078125 */ > + introduce<UQ<1, 7>>("UQ1.7"); > + fails += quantizedCheck<UQ<1, 7>>(-1.0f, 0b0'0000000, 0.0f); /* Clamped to Min */ > + fails += quantizedCheck<UQ<1, 7>>( 0.0f, 0b0'0000000, 0.0f); /* Min / Zero */ > + fails += quantizedCheck<UQ<1, 7>>( 1.0f, 0b1'0000000, 1.0f); /* Mid */ > + fails += quantizedCheck<UQ<1, 7>>( 1.992f, 0b1'1111111, 1.99219f); /* Max */ > + fails += quantizedCheck<UQ<1, 7>>( 2.000f, 0b1'1111111, 1.99219f); /* Clamped to Max */ > + > + /* UQ4.8(0 .. 15.9961) Min: [0x0000:0] -- Max: [0x0fff:15.9961] Step:0.00390625 */ > + introduce<UQ<4, 8>>("UQ4.8"); > + fails += quantizedCheck<UQ<4, 8>>( 0.0f, 0b0000'00000000, 0.00f); > + fails += quantizedCheck<UQ<4, 8>>(16.0f, 0b1111'11111111, 15.9961f); > + > + /* Q5.4(-16 .. 15.9375) Min: [0x0100:-16] -- Max: [0x00ff:15.9375] Step:0.0625 */ > + introduce<Q<5, 4>>("Q5.4"); > + fails += quantizedCheck<Q<5, 4>>(-16.00f, 0b10000'0000, -16.00f); > + fails += quantizedCheck<Q<5, 4>>( 15.94f, 0b01111'1111, 15.9375f); > + > + /* UQ5.8(0 .. 31.9961) Min: [0x0000:0] -- Max: [0x1fff:31.9961] Step:0.00390625 */ > + introduce<UQ<5, 8>>("UQ5.8"); > + fails += quantizedCheck<UQ<5, 8>>( 0.00f, 0b00000'00000000, 0.00f); > + fails += quantizedCheck<UQ<5, 8>>(32.00f, 0b11111'11111111, 31.9961f); > + > + /* Q12.4(-2048 .. 2047.94) Min: [0x8000:-2048] -- Max: [0x7fff:2047.94] Step:0.0625 */ > + introduce<Q<12, 4>>("Q12.4"); > + fails += quantizedCheck<Q<12, 4>>( 0.0f, 0b000000000000'0000, 0.0f); > + fails += quantizedCheck<Q<12, 4>>( 7.5f, 0b000000000111'1000, 7.5f); > + > + /* UQ12.4(0 .. 4095.94) Min: [0x0000:0] -- Max: [0xffff:4095.94] Step:0.0625 */ > + introduce<UQ<12, 4>>("UQ12.4"); > + fails += quantizedCheck<UQ<12, 4>>(0.0f, 0b000000000000'0000, 0.0f); > + fails += quantizedCheck<UQ<12, 4>>(7.5f, 0b000000000111'1000, 7.5f); > + > + /* Q4.20 */ > + introduce<Q<4, 20>>("Q4.20"); > + fails += quantizedCheck<Q<4, 20>>( -9.0f, 0b1000'00000000000000000000, -8.0f); > + fails += quantizedCheck<Q<4, 20>>( -8.0f, 0b1000'00000000000000000000, -8.0f); > + fails += quantizedCheck<Q<4, 20>>( 8.0f, 0b0111'11111111111111111111, 8.0f); > + fails += quantizedCheck<Q<4, 20>>( 9.0f, 0b0111'11111111111111111111, 8.0f); > + > + /* UQ4.20 */ > + introduce<UQ<4, 20>>("UQ4.20"); > + fails += quantizedCheck<UQ<4, 20>>(-1.0f, 0b0000'00000000000000000000, 0.0f); > + fails += quantizedCheck<UQ<4, 20>>( 0.0f, 0b0000'00000000000000000000, 0.0f); > + fails += quantizedCheck<UQ<4, 20>>(16.0f, 0b1111'11111111111111111111, 16.0f); > + fails += quantizedCheck<UQ<4, 20>>(20.0f, 0b1111'11111111111111111111, 16.0f); > + > + /* Storage selection tests */ > + static_assert(std::is_same_v<typename Q<4, 4>::TraitsType::QuantizedType, uint8_t>); > + static_assert(std::is_same_v<typename UQ<4, 4>::TraitsType::QuantizedType, uint8_t>); > + static_assert(std::is_same_v<typename Q<8, 8>::TraitsType::QuantizedType, uint16_t>); > + static_assert(std::is_same_v<typename UQ<8, 8>::TraitsType::QuantizedType, uint16_t>); > + static_assert(std::is_same_v<typename Q<20, 4>::TraitsType::QuantizedType, uint32_t>); > + static_assert(std::is_same_v<typename UQ<20, 4>::TraitsType::QuantizedType, uint32_t>); > + > + /* > + * Test and validate that sign extension can not modify a > + * quantized value when stored or cast to a larger register. > + */ > + std::cout << std::endl; > + std::cout << "Testing sign extension of quantized values when cast to larger registers" << std::endl; > + fails += signExtendCheck<Q<2, 4>, uint8_t>(); > + fails += signExtendCheck<Q<4, 4>, uint8_t>(); > + fails += signExtendCheck<Q<4, 4>, uint16_t>(); > + fails += signExtendCheck<Q<8, 8>, uint16_t>(); > + fails += signExtendCheck<Q<8, 8>, uint32_t>(); > + > + /* clang-format on */ > + > + std::cout << std::endl; > + > + if (fails > 0) { > + cout << "Fixed point quantizer tests failed: " > + << std::dec << fails << " failures." << std::endl; > return TestFail; > + } > > return TestPass; > } > + > + int run() > + { > + unsigned int fails = 0; > + > + /* fixed point conversion test */ > + if (testFixedPoint() != TestPass) > + fails++; > + > + if (testFixedPointQuantizers() != TestPass) > + fails++; > + > + return fails ? TestFail : TestPass; > + } > }; > > TEST_REGISTER(FixedPointUtilsTest) > > -- > 2.52.0 >
Hi Kieran, Thank you for the patch! Quoting Kieran Bingham (2026-02-13 16:57:44) > Provide tests to validate the conditions of FixedPoint types used > within libcamera explicitly. > > Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com> > > --- > v3: > - Rename quantized_type to QuantizedType > > v5: > > - Use string_view for introduction of type > - move min/max check to static assert > - Squash down all tests into a single implementation and remove extra > type proliferation. > - Use Q/UQ types directly. > - use std::cout consistently > > v6: > - Add notes on test introduction > - Expand Q12.4 to include clamp tests to validate int16_t range usage > - Add storage selection assertion checks > - Move clamp checks to their types to expand > - Add {U,}Q<4,20> to check up to 24 bit precision > > v7: > - Provide signed register extension tests > --- > test/ipa/libipa/fixedpoint.cpp | 209 ++++++++++++++++++++++++++++++++++++++++- > 1 file changed, 205 insertions(+), 4 deletions(-) > > diff --git a/test/ipa/libipa/fixedpoint.cpp b/test/ipa/libipa/fixedpoint.cpp > index 4b017e86a74f..2907426b7a69 100644 > --- a/test/ipa/libipa/fixedpoint.cpp > +++ b/test/ipa/libipa/fixedpoint.cpp > @@ -5,12 +5,14 @@ > * Fixed / Floating point utility tests > */ > > +#include "../src/ipa/libipa/fixedpoint.h" > + > #include <cmath> > #include <iostream> > #include <map> > #include <stdint.h> > > -#include "../src/ipa/libipa/fixedpoint.h" > +#include <libcamera/base/utils.h> > > #include "test.h" > > @@ -95,14 +97,213 @@ protected: > return TestPass; > } > > - int run() > + template<typename Q> > + int quantizedCheck(float input, typename Q::QuantizedType expected, float value) > { > - /* fixed point conversion test */ > - if (testFixedPoint() != TestPass) > + Q q(input); > + using T = typename Q::QuantizedType; > + > + std::cout << " Checking " << input << " == " << q << std::endl; > + > + T quantized = q.quantized(); > + if (quantized != expected) { > + std::cout << " ** Q Expected " << input > + << " to quantize to " << utils::hex(expected) > + << ", got " << utils::hex(quantized) > + << " - (" << q << ")" > + << std::endl; > + return 1; > + } > + > + if ((std::abs(q.value() - value)) > 0.0001f) { > + std::cout << " ** V Expected " << input > + << " to quantize to " << value > + << ", got " << q.value() > + << " - (" << q << ")" > + << std::endl; > + return 1; > + } > + > + return 0; > + } > + > + template<typename Q, typename RegisterType> > + int signExtendCheck() > + { > + Q minValue(Q::TraitsType::min); > + using T = typename Q::QuantizedType; > + T minQuantized = minValue.quantized(); > + > + /* Take the value and expand it into a larger register */ > + RegisterType reg = static_cast<RegisterType>(minValue.quantized()); > + > + std::cout << " Checking " << minValue << " == " << utils::hex(reg) << std::endl; > + > + /* > + * Ensure that the minimum negative value is not sign > + * extended when casting up to a larger register storage > + * type. > + */ > + if (reg != minQuantized) { > + std::cout << " ** Sign extension check failed for min value. Expected " > + << utils::hex(minQuantized) << ", got " > + << utils::hex(reg) << std::endl; > + return 1; > + } > + > + /* And we should be consistent when we convert back to Q */ > + Q q2(static_cast<T>(reg)); > + if (!(q2 == minValue)) { > + std::cout << " ** Sign extension check failed for min value. Expected " > + << minValue << ", got " << q2 << std::endl; > + return 1; > + } > + > + return 0; > + } > + > + template<typename Q> > + void introduce(std::string_view type) > + { > + using T = typename Q::QuantizedType; > + > + std::cout << std::endl; > + > + std::cout << type > + << "(" << Q::TraitsType::min << " .. " << Q::TraitsType::max << ") " > + << " Min: " << Q(Q::TraitsType::min) > + << " -- Max: " << Q(Q::TraitsType::max) > + << " Step:" << Q(T(1)).value() > + << std::endl; > + } > + > + int testFixedPointQuantizers() > + { > + unsigned int fails = 0; > + > + /* > + * These aim to specifically test all the corner cases of the > + * quantization and de-quantization process. Including clamping > + * to min/max, zero points and making sure that steps are > + * correct. > + * > + * In particular test signed and unsigned types and a mix of the > + * highest bit width of a storage type and smaller widths that > + * require bit masking and sign extension. > + * > + * Note we must hard code these. Any calculation of expected > + * values risks replicating bugs in the implementation. > + * > + * As the underlying types are integer and float the limit of > + * precision is around 24 bits so we do not test wider types. > + */ > + > + /* clang-format off */ > + > + /* Q1.7(-1 .. 0.992188) Min: [0x80:-1] -- Max: [0x7f:0.992188] Step:0.0078125*/ > + introduce<Q<1, 7>>("Q1.7"); > + fails += quantizedCheck<Q<1, 7>>(-2.000f, 0b1'0000000, -1.0f); /* Clamped to Min */ > + fails += quantizedCheck<Q<1, 7>>(-1.000f, 0b1'0000000, -1.0f); /* Min */ > + fails += quantizedCheck<Q<1, 7>>(-0.992f, 0b1'0000001, -0.992188f); /* Min + 1 step */ > + fails += quantizedCheck<Q<1, 7>>(-0.006f, 0b1'1111111, -0.0078125f); /* -1 step */ > + fails += quantizedCheck<Q<1, 7>>( 0.000f, 0b0'0000000, 0.0f); /* Zero */ > + fails += quantizedCheck<Q<1, 7>>( 0.008f, 0b0'0000001, 0.0078125f); /* +1 step */ > + fails += quantizedCheck<Q<1, 7>>( 0.992f, 0b0'1111111, 0.992188f); /* Max */ > + fails += quantizedCheck<Q<1, 7>>( 2.000f, 0b0'1111111, 0.992188f); /* Clamped to Max */ > + > + /* UQ1.7(0 .. 1.99219) Min: [0x00:0] -- Max: [0xff:1.99219] Step:0.0078125 */ > + introduce<UQ<1, 7>>("UQ1.7"); > + fails += quantizedCheck<UQ<1, 7>>(-1.0f, 0b0'0000000, 0.0f); /* Clamped to Min */ > + fails += quantizedCheck<UQ<1, 7>>( 0.0f, 0b0'0000000, 0.0f); /* Min / Zero */ > + fails += quantizedCheck<UQ<1, 7>>( 1.0f, 0b1'0000000, 1.0f); /* Mid */ > + fails += quantizedCheck<UQ<1, 7>>( 1.992f, 0b1'1111111, 1.99219f); /* Max */ > + fails += quantizedCheck<UQ<1, 7>>( 2.000f, 0b1'1111111, 1.99219f); /* Clamped to Max */ > + > + /* UQ4.8(0 .. 15.9961) Min: [0x0000:0] -- Max: [0x0fff:15.9961] Step:0.00390625 */ > + introduce<UQ<4, 8>>("UQ4.8"); > + fails += quantizedCheck<UQ<4, 8>>( 0.0f, 0b0000'00000000, 0.00f); > + fails += quantizedCheck<UQ<4, 8>>(16.0f, 0b1111'11111111, 15.9961f); > + > + /* Q5.4(-16 .. 15.9375) Min: [0x0100:-16] -- Max: [0x00ff:15.9375] Step:0.0625 */ > + introduce<Q<5, 4>>("Q5.4"); > + fails += quantizedCheck<Q<5, 4>>(-16.00f, 0b10000'0000, -16.00f); > + fails += quantizedCheck<Q<5, 4>>( 15.94f, 0b01111'1111, 15.9375f); > + > + /* UQ5.8(0 .. 31.9961) Min: [0x0000:0] -- Max: [0x1fff:31.9961] Step:0.00390625 */ > + introduce<UQ<5, 8>>("UQ5.8"); > + fails += quantizedCheck<UQ<5, 8>>( 0.00f, 0b00000'00000000, 0.00f); > + fails += quantizedCheck<UQ<5, 8>>(32.00f, 0b11111'11111111, 31.9961f); > + > + /* Q12.4(-2048 .. 2047.94) Min: [0x8000:-2048] -- Max: [0x7fff:2047.94] Step:0.0625 */ > + introduce<Q<12, 4>>("Q12.4"); > + fails += quantizedCheck<Q<12, 4>>( 0.0f, 0b000000000000'0000, 0.0f); > + fails += quantizedCheck<Q<12, 4>>( 7.5f, 0b000000000111'1000, 7.5f); > + > + /* UQ12.4(0 .. 4095.94) Min: [0x0000:0] -- Max: [0xffff:4095.94] Step:0.0625 */ > + introduce<UQ<12, 4>>("UQ12.4"); > + fails += quantizedCheck<UQ<12, 4>>(0.0f, 0b000000000000'0000, 0.0f); > + fails += quantizedCheck<UQ<12, 4>>(7.5f, 0b000000000111'1000, 7.5f); > + > + /* Q4.20 */ > + introduce<Q<4, 20>>("Q4.20"); > + fails += quantizedCheck<Q<4, 20>>( -9.0f, 0b1000'00000000000000000000, -8.0f); > + fails += quantizedCheck<Q<4, 20>>( -8.0f, 0b1000'00000000000000000000, -8.0f); > + fails += quantizedCheck<Q<4, 20>>( 8.0f, 0b0111'11111111111111111111, 8.0f); > + fails += quantizedCheck<Q<4, 20>>( 9.0f, 0b0111'11111111111111111111, 8.0f); > + > + /* UQ4.20 */ > + introduce<UQ<4, 20>>("UQ4.20"); > + fails += quantizedCheck<UQ<4, 20>>(-1.0f, 0b0000'00000000000000000000, 0.0f); > + fails += quantizedCheck<UQ<4, 20>>( 0.0f, 0b0000'00000000000000000000, 0.0f); > + fails += quantizedCheck<UQ<4, 20>>(16.0f, 0b1111'11111111111111111111, 16.0f); > + fails += quantizedCheck<UQ<4, 20>>(20.0f, 0b1111'11111111111111111111, 16.0f); > + > + /* Storage selection tests */ > + static_assert(std::is_same_v<typename Q<4, 4>::TraitsType::QuantizedType, uint8_t>); > + static_assert(std::is_same_v<typename UQ<4, 4>::TraitsType::QuantizedType, uint8_t>); > + static_assert(std::is_same_v<typename Q<8, 8>::TraitsType::QuantizedType, uint16_t>); > + static_assert(std::is_same_v<typename UQ<8, 8>::TraitsType::QuantizedType, uint16_t>); > + static_assert(std::is_same_v<typename Q<20, 4>::TraitsType::QuantizedType, uint32_t>); > + static_assert(std::is_same_v<typename UQ<20, 4>::TraitsType::QuantizedType, uint32_t>); > + > + /* > + * Test and validate that sign extension can not modify a > + * quantized value when stored or cast to a larger register. > + */ > + std::cout << std::endl; > + std::cout << "Testing sign extension of quantized values when cast to larger registers" << std::endl; > + fails += signExtendCheck<Q<2, 4>, uint8_t>(); > + fails += signExtendCheck<Q<4, 4>, uint8_t>(); > + fails += signExtendCheck<Q<4, 4>, uint16_t>(); > + fails += signExtendCheck<Q<8, 8>, uint16_t>(); > + fails += signExtendCheck<Q<8, 8>, uint32_t>(); > + > + /* clang-format on */ > + > + std::cout << std::endl; > + > + if (fails > 0) { > + cout << "Fixed point quantizer tests failed: " > + << std::dec << fails << " failures." << std::endl; > return TestFail; > + } > > return TestPass; > } > + > + int run() > + { > + unsigned int fails = 0; > + > + /* fixed point conversion test */ > + if (testFixedPoint() != TestPass) > + fails++; > + > + if (testFixedPointQuantizers() != TestPass) > + fails++; > + > + return fails ? TestFail : TestPass; > + } > }; Reviewed-by: Isaac Scott <isaac.scott@ideasonboard.com> > > TEST_REGISTER(FixedPointUtilsTest) > > -- > 2.52.0 >
Hi Kieran, Thank you for the patch. On Fri, Feb 13, 2026 at 04:57:44PM +0000, Kieran Bingham wrote: > Provide tests to validate the conditions of FixedPoint types used > within libcamera explicitly. > > Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com> > > --- > v3: > - Rename quantized_type to QuantizedType > > v5: > > - Use string_view for introduction of type > - move min/max check to static assert > - Squash down all tests into a single implementation and remove extra > type proliferation. > - Use Q/UQ types directly. > - use std::cout consistently > > v6: > - Add notes on test introduction > - Expand Q12.4 to include clamp tests to validate int16_t range usage > - Add storage selection assertion checks > - Move clamp checks to their types to expand > - Add {U,}Q<4,20> to check up to 24 bit precision > > v7: > - Provide signed register extension tests > --- > test/ipa/libipa/fixedpoint.cpp | 209 ++++++++++++++++++++++++++++++++++++++++- > 1 file changed, 205 insertions(+), 4 deletions(-) > > diff --git a/test/ipa/libipa/fixedpoint.cpp b/test/ipa/libipa/fixedpoint.cpp > index 4b017e86a74f..2907426b7a69 100644 > --- a/test/ipa/libipa/fixedpoint.cpp > +++ b/test/ipa/libipa/fixedpoint.cpp > @@ -5,12 +5,14 @@ > * Fixed / Floating point utility tests > */ > > +#include "../src/ipa/libipa/fixedpoint.h" > + > #include <cmath> > #include <iostream> > #include <map> > #include <stdint.h> > > -#include "../src/ipa/libipa/fixedpoint.h" > +#include <libcamera/base/utils.h> > > #include "test.h" > > @@ -95,14 +97,213 @@ protected: > return TestPass; > } > > - int run() > + template<typename Q> > + int quantizedCheck(float input, typename Q::QuantizedType expected, float value) > { > - /* fixed point conversion test */ > - if (testFixedPoint() != TestPass) > + Q q(input); > + using T = typename Q::QuantizedType; > + > + std::cout << " Checking " << input << " == " << q << std::endl; > + > + T quantized = q.quantized(); > + if (quantized != expected) { > + std::cout << " ** Q Expected " << input std::cerr as this is an error message. > + << " to quantize to " << utils::hex(expected) > + << ", got " << utils::hex(quantized) > + << " - (" << q << ")" > + << std::endl; > + return 1; > + } > + > + if ((std::abs(q.value() - value)) > 0.0001f) { > + std::cout << " ** V Expected " << input Same here. > + << " to quantize to " << value > + << ", got " << q.value() > + << " - (" << q << ")" > + << std::endl; > + return 1; > + } > + > + return 0; > + } > + > + template<typename Q, typename RegisterType> > + int signExtendCheck() > + { > + Q minValue(Q::TraitsType::min); > + using T = typename Q::QuantizedType; > + T minQuantized = minValue.quantized(); > + > + /* Take the value and expand it into a larger register */ > + RegisterType reg = static_cast<RegisterType>(minValue.quantized()); > + > + std::cout << " Checking " << minValue << " == " << utils::hex(reg) << std::endl; > + > + /* > + * Ensure that the minimum negative value is not sign > + * extended when casting up to a larger register storage > + * type. > + */ > + if (reg != minQuantized) { > + std::cout << " ** Sign extension check failed for min value. Expected " Here too. > + << utils::hex(minQuantized) << ", got " > + << utils::hex(reg) << std::endl; > + return 1; > + } > + > + /* And we should be consistent when we convert back to Q */ > + Q q2(static_cast<T>(reg)); > + if (!(q2 == minValue)) { > + std::cout << " ** Sign extension check failed for min value. Expected " And here. > + << minValue << ", got " << q2 << std::endl; > + return 1; > + } > + > + return 0; > + } > + > + template<typename Q> > + void introduce(std::string_view type) > + { > + using T = typename Q::QuantizedType; > + > + std::cout << std::endl; > + > + std::cout << type > + << "(" << Q::TraitsType::min << " .. " << Q::TraitsType::max << ") " > + << " Min: " << Q(Q::TraitsType::min) > + << " -- Max: " << Q(Q::TraitsType::max) > + << " Step:" << Q(T(1)).value() > + << std::endl; > + } > + > + int testFixedPointQuantizers() > + { > + unsigned int fails = 0; > + > + /* > + * These aim to specifically test all the corner cases of the > + * quantization and de-quantization process. Including clamping > + * to min/max, zero points and making sure that steps are > + * correct. > + * > + * In particular test signed and unsigned types and a mix of the > + * highest bit width of a storage type and smaller widths that > + * require bit masking and sign extension. > + * > + * Note we must hard code these. Any calculation of expected > + * values risks replicating bugs in the implementation. > + * > + * As the underlying types are integer and float the limit of > + * precision is around 24 bits so we do not test wider types. > + */ > + > + /* clang-format off */ > + > + /* Q1.7(-1 .. 0.992188) Min: [0x80:-1] -- Max: [0x7f:0.992188] Step:0.0078125*/ > + introduce<Q<1, 7>>("Q1.7"); > + fails += quantizedCheck<Q<1, 7>>(-2.000f, 0b1'0000000, -1.0f); /* Clamped to Min */ > + fails += quantizedCheck<Q<1, 7>>(-1.000f, 0b1'0000000, -1.0f); /* Min */ > + fails += quantizedCheck<Q<1, 7>>(-0.992f, 0b1'0000001, -0.992188f); /* Min + 1 step */ > + fails += quantizedCheck<Q<1, 7>>(-0.006f, 0b1'1111111, -0.0078125f); /* -1 step */ > + fails += quantizedCheck<Q<1, 7>>( 0.000f, 0b0'0000000, 0.0f); /* Zero */ > + fails += quantizedCheck<Q<1, 7>>( 0.008f, 0b0'0000001, 0.0078125f); /* +1 step */ > + fails += quantizedCheck<Q<1, 7>>( 0.992f, 0b0'1111111, 0.992188f); /* Max */ > + fails += quantizedCheck<Q<1, 7>>( 2.000f, 0b0'1111111, 0.992188f); /* Clamped to Max */ > + > + /* UQ1.7(0 .. 1.99219) Min: [0x00:0] -- Max: [0xff:1.99219] Step:0.0078125 */ > + introduce<UQ<1, 7>>("UQ1.7"); > + fails += quantizedCheck<UQ<1, 7>>(-1.0f, 0b0'0000000, 0.0f); /* Clamped to Min */ > + fails += quantizedCheck<UQ<1, 7>>( 0.0f, 0b0'0000000, 0.0f); /* Min / Zero */ > + fails += quantizedCheck<UQ<1, 7>>( 1.0f, 0b1'0000000, 1.0f); /* Mid */ > + fails += quantizedCheck<UQ<1, 7>>( 1.992f, 0b1'1111111, 1.99219f); /* Max */ > + fails += quantizedCheck<UQ<1, 7>>( 2.000f, 0b1'1111111, 1.99219f); /* Clamped to Max */ > + > + /* UQ4.8(0 .. 15.9961) Min: [0x0000:0] -- Max: [0x0fff:15.9961] Step:0.00390625 */ > + introduce<UQ<4, 8>>("UQ4.8"); > + fails += quantizedCheck<UQ<4, 8>>( 0.0f, 0b0000'00000000, 0.00f); > + fails += quantizedCheck<UQ<4, 8>>(16.0f, 0b1111'11111111, 15.9961f); > + > + /* Q5.4(-16 .. 15.9375) Min: [0x0100:-16] -- Max: [0x00ff:15.9375] Step:0.0625 */ > + introduce<Q<5, 4>>("Q5.4"); > + fails += quantizedCheck<Q<5, 4>>(-16.00f, 0b10000'0000, -16.00f); > + fails += quantizedCheck<Q<5, 4>>( 15.94f, 0b01111'1111, 15.9375f); > + > + /* UQ5.8(0 .. 31.9961) Min: [0x0000:0] -- Max: [0x1fff:31.9961] Step:0.00390625 */ > + introduce<UQ<5, 8>>("UQ5.8"); > + fails += quantizedCheck<UQ<5, 8>>( 0.00f, 0b00000'00000000, 0.00f); > + fails += quantizedCheck<UQ<5, 8>>(32.00f, 0b11111'11111111, 31.9961f); > + > + /* Q12.4(-2048 .. 2047.94) Min: [0x8000:-2048] -- Max: [0x7fff:2047.94] Step:0.0625 */ > + introduce<Q<12, 4>>("Q12.4"); > + fails += quantizedCheck<Q<12, 4>>( 0.0f, 0b000000000000'0000, 0.0f); > + fails += quantizedCheck<Q<12, 4>>( 7.5f, 0b000000000111'1000, 7.5f); > + > + /* UQ12.4(0 .. 4095.94) Min: [0x0000:0] -- Max: [0xffff:4095.94] Step:0.0625 */ > + introduce<UQ<12, 4>>("UQ12.4"); > + fails += quantizedCheck<UQ<12, 4>>(0.0f, 0b000000000000'0000, 0.0f); > + fails += quantizedCheck<UQ<12, 4>>(7.5f, 0b000000000111'1000, 7.5f); > + > + /* Q4.20 */ > + introduce<Q<4, 20>>("Q4.20"); > + fails += quantizedCheck<Q<4, 20>>( -9.0f, 0b1000'00000000000000000000, -8.0f); > + fails += quantizedCheck<Q<4, 20>>( -8.0f, 0b1000'00000000000000000000, -8.0f); > + fails += quantizedCheck<Q<4, 20>>( 8.0f, 0b0111'11111111111111111111, 8.0f); > + fails += quantizedCheck<Q<4, 20>>( 9.0f, 0b0111'11111111111111111111, 8.0f); > + > + /* UQ4.20 */ > + introduce<UQ<4, 20>>("UQ4.20"); > + fails += quantizedCheck<UQ<4, 20>>(-1.0f, 0b0000'00000000000000000000, 0.0f); > + fails += quantizedCheck<UQ<4, 20>>( 0.0f, 0b0000'00000000000000000000, 0.0f); > + fails += quantizedCheck<UQ<4, 20>>(16.0f, 0b1111'11111111111111111111, 16.0f); > + fails += quantizedCheck<UQ<4, 20>>(20.0f, 0b1111'11111111111111111111, 16.0f); > + > + /* Storage selection tests */ > + static_assert(std::is_same_v<typename Q<4, 4>::TraitsType::QuantizedType, uint8_t>); > + static_assert(std::is_same_v<typename UQ<4, 4>::TraitsType::QuantizedType, uint8_t>); > + static_assert(std::is_same_v<typename Q<8, 8>::TraitsType::QuantizedType, uint16_t>); > + static_assert(std::is_same_v<typename UQ<8, 8>::TraitsType::QuantizedType, uint16_t>); > + static_assert(std::is_same_v<typename Q<20, 4>::TraitsType::QuantizedType, uint32_t>); > + static_assert(std::is_same_v<typename UQ<20, 4>::TraitsType::QuantizedType, uint32_t>); > + > + /* > + * Test and validate that sign extension can not modify a > + * quantized value when stored or cast to a larger register. > + */ > + std::cout << std::endl; > + std::cout << "Testing sign extension of quantized values when cast to larger registers" << std::endl; > + fails += signExtendCheck<Q<2, 4>, uint8_t>(); > + fails += signExtendCheck<Q<4, 4>, uint8_t>(); > + fails += signExtendCheck<Q<4, 4>, uint16_t>(); > + fails += signExtendCheck<Q<8, 8>, uint16_t>(); > + fails += signExtendCheck<Q<8, 8>, uint32_t>(); > + > + /* clang-format on */ > + > + std::cout << std::endl; > + > + if (fails > 0) { > + cout << "Fixed point quantizer tests failed: " s/cout/std::cerr/ Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com> > + << std::dec << fails << " failures." << std::endl; > return TestFail; > + } > > return TestPass; > } > + > + int run() > + { > + unsigned int fails = 0; > + > + /* fixed point conversion test */ > + if (testFixedPoint() != TestPass) > + fails++; > + > + if (testFixedPointQuantizers() != TestPass) > + fails++; > + > + return fails ? TestFail : TestPass; > + } > }; > > TEST_REGISTER(FixedPointUtilsTest)
diff --git a/test/ipa/libipa/fixedpoint.cpp b/test/ipa/libipa/fixedpoint.cpp index 4b017e86a74f..2907426b7a69 100644 --- a/test/ipa/libipa/fixedpoint.cpp +++ b/test/ipa/libipa/fixedpoint.cpp @@ -5,12 +5,14 @@ * Fixed / Floating point utility tests */ +#include "../src/ipa/libipa/fixedpoint.h" + #include <cmath> #include <iostream> #include <map> #include <stdint.h> -#include "../src/ipa/libipa/fixedpoint.h" +#include <libcamera/base/utils.h> #include "test.h" @@ -95,14 +97,213 @@ protected: return TestPass; } - int run() + template<typename Q> + int quantizedCheck(float input, typename Q::QuantizedType expected, float value) { - /* fixed point conversion test */ - if (testFixedPoint() != TestPass) + Q q(input); + using T = typename Q::QuantizedType; + + std::cout << " Checking " << input << " == " << q << std::endl; + + T quantized = q.quantized(); + if (quantized != expected) { + std::cout << " ** Q Expected " << input + << " to quantize to " << utils::hex(expected) + << ", got " << utils::hex(quantized) + << " - (" << q << ")" + << std::endl; + return 1; + } + + if ((std::abs(q.value() - value)) > 0.0001f) { + std::cout << " ** V Expected " << input + << " to quantize to " << value + << ", got " << q.value() + << " - (" << q << ")" + << std::endl; + return 1; + } + + return 0; + } + + template<typename Q, typename RegisterType> + int signExtendCheck() + { + Q minValue(Q::TraitsType::min); + using T = typename Q::QuantizedType; + T minQuantized = minValue.quantized(); + + /* Take the value and expand it into a larger register */ + RegisterType reg = static_cast<RegisterType>(minValue.quantized()); + + std::cout << " Checking " << minValue << " == " << utils::hex(reg) << std::endl; + + /* + * Ensure that the minimum negative value is not sign + * extended when casting up to a larger register storage + * type. + */ + if (reg != minQuantized) { + std::cout << " ** Sign extension check failed for min value. Expected " + << utils::hex(minQuantized) << ", got " + << utils::hex(reg) << std::endl; + return 1; + } + + /* And we should be consistent when we convert back to Q */ + Q q2(static_cast<T>(reg)); + if (!(q2 == minValue)) { + std::cout << " ** Sign extension check failed for min value. Expected " + << minValue << ", got " << q2 << std::endl; + return 1; + } + + return 0; + } + + template<typename Q> + void introduce(std::string_view type) + { + using T = typename Q::QuantizedType; + + std::cout << std::endl; + + std::cout << type + << "(" << Q::TraitsType::min << " .. " << Q::TraitsType::max << ") " + << " Min: " << Q(Q::TraitsType::min) + << " -- Max: " << Q(Q::TraitsType::max) + << " Step:" << Q(T(1)).value() + << std::endl; + } + + int testFixedPointQuantizers() + { + unsigned int fails = 0; + + /* + * These aim to specifically test all the corner cases of the + * quantization and de-quantization process. Including clamping + * to min/max, zero points and making sure that steps are + * correct. + * + * In particular test signed and unsigned types and a mix of the + * highest bit width of a storage type and smaller widths that + * require bit masking and sign extension. + * + * Note we must hard code these. Any calculation of expected + * values risks replicating bugs in the implementation. + * + * As the underlying types are integer and float the limit of + * precision is around 24 bits so we do not test wider types. + */ + + /* clang-format off */ + + /* Q1.7(-1 .. 0.992188) Min: [0x80:-1] -- Max: [0x7f:0.992188] Step:0.0078125*/ + introduce<Q<1, 7>>("Q1.7"); + fails += quantizedCheck<Q<1, 7>>(-2.000f, 0b1'0000000, -1.0f); /* Clamped to Min */ + fails += quantizedCheck<Q<1, 7>>(-1.000f, 0b1'0000000, -1.0f); /* Min */ + fails += quantizedCheck<Q<1, 7>>(-0.992f, 0b1'0000001, -0.992188f); /* Min + 1 step */ + fails += quantizedCheck<Q<1, 7>>(-0.006f, 0b1'1111111, -0.0078125f); /* -1 step */ + fails += quantizedCheck<Q<1, 7>>( 0.000f, 0b0'0000000, 0.0f); /* Zero */ + fails += quantizedCheck<Q<1, 7>>( 0.008f, 0b0'0000001, 0.0078125f); /* +1 step */ + fails += quantizedCheck<Q<1, 7>>( 0.992f, 0b0'1111111, 0.992188f); /* Max */ + fails += quantizedCheck<Q<1, 7>>( 2.000f, 0b0'1111111, 0.992188f); /* Clamped to Max */ + + /* UQ1.7(0 .. 1.99219) Min: [0x00:0] -- Max: [0xff:1.99219] Step:0.0078125 */ + introduce<UQ<1, 7>>("UQ1.7"); + fails += quantizedCheck<UQ<1, 7>>(-1.0f, 0b0'0000000, 0.0f); /* Clamped to Min */ + fails += quantizedCheck<UQ<1, 7>>( 0.0f, 0b0'0000000, 0.0f); /* Min / Zero */ + fails += quantizedCheck<UQ<1, 7>>( 1.0f, 0b1'0000000, 1.0f); /* Mid */ + fails += quantizedCheck<UQ<1, 7>>( 1.992f, 0b1'1111111, 1.99219f); /* Max */ + fails += quantizedCheck<UQ<1, 7>>( 2.000f, 0b1'1111111, 1.99219f); /* Clamped to Max */ + + /* UQ4.8(0 .. 15.9961) Min: [0x0000:0] -- Max: [0x0fff:15.9961] Step:0.00390625 */ + introduce<UQ<4, 8>>("UQ4.8"); + fails += quantizedCheck<UQ<4, 8>>( 0.0f, 0b0000'00000000, 0.00f); + fails += quantizedCheck<UQ<4, 8>>(16.0f, 0b1111'11111111, 15.9961f); + + /* Q5.4(-16 .. 15.9375) Min: [0x0100:-16] -- Max: [0x00ff:15.9375] Step:0.0625 */ + introduce<Q<5, 4>>("Q5.4"); + fails += quantizedCheck<Q<5, 4>>(-16.00f, 0b10000'0000, -16.00f); + fails += quantizedCheck<Q<5, 4>>( 15.94f, 0b01111'1111, 15.9375f); + + /* UQ5.8(0 .. 31.9961) Min: [0x0000:0] -- Max: [0x1fff:31.9961] Step:0.00390625 */ + introduce<UQ<5, 8>>("UQ5.8"); + fails += quantizedCheck<UQ<5, 8>>( 0.00f, 0b00000'00000000, 0.00f); + fails += quantizedCheck<UQ<5, 8>>(32.00f, 0b11111'11111111, 31.9961f); + + /* Q12.4(-2048 .. 2047.94) Min: [0x8000:-2048] -- Max: [0x7fff:2047.94] Step:0.0625 */ + introduce<Q<12, 4>>("Q12.4"); + fails += quantizedCheck<Q<12, 4>>( 0.0f, 0b000000000000'0000, 0.0f); + fails += quantizedCheck<Q<12, 4>>( 7.5f, 0b000000000111'1000, 7.5f); + + /* UQ12.4(0 .. 4095.94) Min: [0x0000:0] -- Max: [0xffff:4095.94] Step:0.0625 */ + introduce<UQ<12, 4>>("UQ12.4"); + fails += quantizedCheck<UQ<12, 4>>(0.0f, 0b000000000000'0000, 0.0f); + fails += quantizedCheck<UQ<12, 4>>(7.5f, 0b000000000111'1000, 7.5f); + + /* Q4.20 */ + introduce<Q<4, 20>>("Q4.20"); + fails += quantizedCheck<Q<4, 20>>( -9.0f, 0b1000'00000000000000000000, -8.0f); + fails += quantizedCheck<Q<4, 20>>( -8.0f, 0b1000'00000000000000000000, -8.0f); + fails += quantizedCheck<Q<4, 20>>( 8.0f, 0b0111'11111111111111111111, 8.0f); + fails += quantizedCheck<Q<4, 20>>( 9.0f, 0b0111'11111111111111111111, 8.0f); + + /* UQ4.20 */ + introduce<UQ<4, 20>>("UQ4.20"); + fails += quantizedCheck<UQ<4, 20>>(-1.0f, 0b0000'00000000000000000000, 0.0f); + fails += quantizedCheck<UQ<4, 20>>( 0.0f, 0b0000'00000000000000000000, 0.0f); + fails += quantizedCheck<UQ<4, 20>>(16.0f, 0b1111'11111111111111111111, 16.0f); + fails += quantizedCheck<UQ<4, 20>>(20.0f, 0b1111'11111111111111111111, 16.0f); + + /* Storage selection tests */ + static_assert(std::is_same_v<typename Q<4, 4>::TraitsType::QuantizedType, uint8_t>); + static_assert(std::is_same_v<typename UQ<4, 4>::TraitsType::QuantizedType, uint8_t>); + static_assert(std::is_same_v<typename Q<8, 8>::TraitsType::QuantizedType, uint16_t>); + static_assert(std::is_same_v<typename UQ<8, 8>::TraitsType::QuantizedType, uint16_t>); + static_assert(std::is_same_v<typename Q<20, 4>::TraitsType::QuantizedType, uint32_t>); + static_assert(std::is_same_v<typename UQ<20, 4>::TraitsType::QuantizedType, uint32_t>); + + /* + * Test and validate that sign extension can not modify a + * quantized value when stored or cast to a larger register. + */ + std::cout << std::endl; + std::cout << "Testing sign extension of quantized values when cast to larger registers" << std::endl; + fails += signExtendCheck<Q<2, 4>, uint8_t>(); + fails += signExtendCheck<Q<4, 4>, uint8_t>(); + fails += signExtendCheck<Q<4, 4>, uint16_t>(); + fails += signExtendCheck<Q<8, 8>, uint16_t>(); + fails += signExtendCheck<Q<8, 8>, uint32_t>(); + + /* clang-format on */ + + std::cout << std::endl; + + if (fails > 0) { + cout << "Fixed point quantizer tests failed: " + << std::dec << fails << " failures." << std::endl; return TestFail; + } return TestPass; } + + int run() + { + unsigned int fails = 0; + + /* fixed point conversion test */ + if (testFixedPoint() != TestPass) + fails++; + + if (testFixedPointQuantizers() != TestPass) + fails++; + + return fails ? TestFail : TestPass; + } }; TEST_REGISTER(FixedPointUtilsTest)
Provide tests to validate the conditions of FixedPoint types used within libcamera explicitly. Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com> --- v3: - Rename quantized_type to QuantizedType v5: - Use string_view for introduction of type - move min/max check to static assert - Squash down all tests into a single implementation and remove extra type proliferation. - Use Q/UQ types directly. - use std::cout consistently v6: - Add notes on test introduction - Expand Q12.4 to include clamp tests to validate int16_t range usage - Add storage selection assertion checks - Move clamp checks to their types to expand - Add {U,}Q<4,20> to check up to 24 bit precision v7: - Provide signed register extension tests --- test/ipa/libipa/fixedpoint.cpp | 209 ++++++++++++++++++++++++++++++++++++++++- 1 file changed, 205 insertions(+), 4 deletions(-)