#include "FastMathF16.h" #include #include "Error.h" #include "Test.h" #define SNR_THRESHOLD 60 #define SNR_LOG_THRESHOLD 40 #define SNR_ATAN2_THRESHOLD 60 /* Reference patterns are generated with a double precision computation. */ #define REL_ERROR (1.0e-3) #define ABS_ERROR (1.0e-3) #define REL_ERROR_ATAN (1.0e-3) #define ABS_ERROR_ATAN (2.0e-3) #if 0 void FastMathF16::test_cos_f16() { const float16_t *inp = input.ptr(); float16_t *outp = output.ptr(); unsigned long i; for(i=0; i < ref.nbSamples(); i++) { outp[i]=arm_cos_f16(inp[i]); } ASSERT_SNR(ref,output,(float16_t)SNR_THRESHOLD); ASSERT_CLOSE_ERROR(ref,output,ABS_ERROR,REL_ERROR); } void FastMathF16::test_sin_f16() { const float16_t *inp = input.ptr(); float16_t *outp = output.ptr(); unsigned long i; for(i=0; i < ref.nbSamples(); i++) { outp[i]=arm_sin_f16(inp[i]); } ASSERT_SNR(ref,output,(float16_t)SNR_THRESHOLD); ASSERT_CLOSE_ERROR(ref,output,ABS_ERROR,REL_ERROR); } #endif void FastMathF16::test_atan2_scalar_f16() { const float16_t *inp = input.ptr(); float16_t *outp = output.ptr(); float16_t res; unsigned long i; arm_status status=ARM_MATH_SUCCESS; for(i=0; i < ref.nbSamples(); i++) { status=arm_atan2_f16(inp[2*i],inp[2*i+1],&res); outp[i]=res; ASSERT_TRUE((status == ARM_MATH_SUCCESS)); } //printf("%f %f %f\n",inp[2*i],inp[2*i+1],outp[i]); ASSERT_SNR(ref,output,(float16_t)SNR_ATAN2_THRESHOLD); ASSERT_CLOSE_ERROR(ref,output,ABS_ERROR_ATAN,REL_ERROR_ATAN); } void FastMathF16::test_sqrt_f16() { const float16_t *inp = input.ptr(); float16_t *outp = output.ptr(); arm_status status; unsigned long i; for(i=0; i < ref.nbSamples(); i++) { status=arm_sqrt_f16(inp[i],&outp[i]); ASSERT_TRUE((status == ARM_MATH_SUCCESS) || ((inp[i] < 0.0f) && (status == ARM_MATH_ARGUMENT_ERROR))); } ASSERT_SNR(ref,output,(float16_t)SNR_THRESHOLD); ASSERT_CLOSE_ERROR(ref,output,ABS_ERROR,REL_ERROR); } void FastMathF16::test_vlog_f16() { const float16_t *inp = input.ptr(); float16_t *outp = output.ptr(); arm_vlog_f16(inp,outp,ref.nbSamples()); //ASSERT_SNR(ref,output,(float16_t)SNR_THRESHOLD); ASSERT_CLOSE_ERROR(ref,output,ABS_ERROR,REL_ERROR); ASSERT_EMPTY_TAIL(output); } void FastMathF16::test_vexp_f16() { const float16_t *inp = input.ptr(); float16_t *outp = output.ptr(); arm_vexp_f16(inp,outp,ref.nbSamples()); ASSERT_CLOSE_ERROR(ref,output,ABS_ERROR,REL_ERROR); ASSERT_SNR(ref,output,(float16_t)SNR_THRESHOLD); ASSERT_EMPTY_TAIL(output); } void FastMathF16::test_inverse_f16() { const float16_t *inp = input.ptr(); float16_t *outp = output.ptr(); arm_vinverse_f16(inp,outp,ref.nbSamples()); ASSERT_CLOSE_ERROR(ref,output,ABS_ERROR,REL_ERROR); ASSERT_SNR(ref,output,(float16_t)SNR_THRESHOLD); ASSERT_EMPTY_TAIL(output); } void FastMathF16::setUp(Testing::testID_t id,std::vector& paramsArgs,Client::PatternMgr *mgr) { (void)paramsArgs; switch(id) { #if 0 case FastMathF16::TEST_COS_F16_1: { input.reload(FastMathF16::ANGLES1_F16_ID,mgr); ref.reload(FastMathF16::COS1_F16_ID,mgr); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_SIN_F16_2: { input.reload(FastMathF16::ANGLES1_F16_ID,mgr); ref.reload(FastMathF16::SIN1_F16_ID,mgr); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; #endif case FastMathF16::TEST_SQRT_F16_3: { input.reload(FastMathF16::SQRTINPUT1_F16_ID,mgr); ref.reload(FastMathF16::SQRT1_F16_ID,mgr); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_VLOG_F16_4: { input.reload(FastMathF16::LOGINPUT1_F16_ID,mgr); ref.reload(FastMathF16::LOG1_F16_ID,mgr); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_VLOG_F16_5: { input.reload(FastMathF16::LOGINPUT1_F16_ID,mgr,7); ref.reload(FastMathF16::LOG1_F16_ID,mgr,7); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_VLOG_F16_6: { input.reload(FastMathF16::LOGINPUT1_F16_ID,mgr,16); ref.reload(FastMathF16::LOG1_F16_ID,mgr,16); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_VLOG_F16_7: { input.reload(FastMathF16::LOGINPUT1_F16_ID,mgr,23); ref.reload(FastMathF16::LOG1_F16_ID,mgr,23); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_VEXP_F16_8: { input.reload(FastMathF16::EXPINPUT1_F16_ID,mgr); ref.reload(FastMathF16::EXP1_F16_ID,mgr); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_VEXP_F16_9: { input.reload(FastMathF16::EXPINPUT1_F16_ID,mgr,7); ref.reload(FastMathF16::EXP1_F16_ID,mgr,7); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_VEXP_F16_10: { input.reload(FastMathF16::EXPINPUT1_F16_ID,mgr,16); ref.reload(FastMathF16::EXP1_F16_ID,mgr,16); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_VEXP_F16_11: { input.reload(FastMathF16::EXPINPUT1_F16_ID,mgr,23); ref.reload(FastMathF16::EXP1_F16_ID,mgr,23); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_INVERSE_F16_12: { input.reload(FastMathF16::INPUT1_F16_ID,mgr); ref.reload(FastMathF16::INVERSE1_F16_ID,mgr); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; case FastMathF16::TEST_ATAN2_SCALAR_F16_13: { input.reload(FastMathF16::ATAN2INPUT1_F16_ID,mgr); ref.reload(FastMathF16::ATAN2_F16_ID,mgr); output.create(ref.nbSamples(),FastMathF16::OUT_F16_ID,mgr); } break; } } void FastMathF16::tearDown(Testing::testID_t id,Client::PatternMgr *mgr) { (void)id; output.dump(mgr); }