diff --git a/include/core/tensor.h b/include/core/tensor.h index 63efd0f7..ff4f3480 100644 --- a/include/core/tensor.h +++ b/include/core/tensor.h @@ -145,14 +145,17 @@ class TensorObj : public TensorBaseObj { void printData() const; bool equalData(const Tensor &rhs, double relativeError = 1e-6) const; - template bool equalData(const vector &dataVector) { + template + bool equalData(const vector &dataVector, double relativeError = 1e-6) { IT_ASSERT(size() == dataVector.size()); if (dtype == DataType::Float16) { return equalDataImpl_fp16(getRawDataPtr(), - (float *)dataVector.data(), size()); + (float *)dataVector.data(), size(), + relativeError); } IT_ASSERT(DataType::get() == dtype.cpuTypeInt()); - return equalDataImpl(getRawDataPtr(), dataVector.data(), size()); + return equalDataImpl(getRawDataPtr(), dataVector.data(), size(), + relativeError); } size_t getOffsetByBroadcastOffset(size_t bcOffset, Shape bcShape) const; @@ -198,24 +201,34 @@ class TensorObj : public TensorBaseObj { if (a[i] != b[i]) return false; } else if constexpr (std::is_floating_point_v) { - if (fabs(a[i] - b[i]) / std::max(fabs(a[i]), fabs(b[i])) > - relativeError) { - printf("Error on %lu: %f %f\n", i, a[i], b[i]); - return false; + if (fabs(b[i]) < 1e-6) { + if (fabs(a[i] - b[i]) > relativeError) { + printf("Error on %lu: %f %f\n", i, a[i], b[i]); + return false; + } + } else { + if (fabs(a[i] - b[i]) / + (std::max(fabs(a[i]), fabs(b[i])) + 1e-6) > + relativeError) { + printf("Error on %lu: %f %f\n", i, a[i], b[i]); + return false; + } } + } else static_assert(!sizeof(T), "Unsupported data type"); } return true; } - bool equalDataImpl_fp16(const uint16_t *a, const float *b, - size_t size) const { + bool equalDataImpl_fp16(const uint16_t *a, const float *b, size_t size, + double relativeError = 1e-6) const { for (size_t i = 0; i < size; ++i) { auto a_fp32 = fp16_to_float(a[i]); auto b_fp32 = b[i]; - if (fabs(a_fp32 - b_fp32) / std::max(fabs(a_fp32), fabs(b_fp32)) > - 1e-6) { + if (fabs(a_fp32 - b_fp32) / + (std::max(fabs(a_fp32), fabs(b_fp32)) + 1e-6) > + relativeError) { printf("Error on %lu: %f %f\n", i, a_fp32, b_fp32); return false; } diff --git a/src/kernels/mlu/include/highSoftmax.h b/src/kernels/mlu/include/highSoftmax.h index 60c3954f..df8d0539 100644 --- a/src/kernels/mlu/include/highSoftmax.h +++ b/src/kernels/mlu/include/highSoftmax.h @@ -1,8 +1,8 @@ -#ifndef BANG_KERNELS_DIVOPERATION_DIV_H_ -#define BANG_KERNELS_DIVOPERATION_DIV_H_ +#ifndef BANG_KERNELS_SOFTMAXOPERATION_SOFTMAX_H_ +#define BANG_KERNELS_SOFTMAXOPERATION_SOFTMAX_H_ __mlu_global__ void softmaxUnion1(float *mlu_destination, float *mlu_src, int nDim, int axis, int othersize, int frontsize, int dimsize, int stride); -#endif // BANG_KERNELS_DIVOPERATION_DIV_H_ +#endif // BANG_KERNELS_SOFTMAXOPERATION_SOFTMAX_H_ diff --git a/src/kernels/mlu/src/highSoftmax_device.mlu b/src/kernels/mlu/src/highSoftmax_device.mlu index b9e47d96..276f69a4 100644 --- a/src/kernels/mlu/src/highSoftmax_device.mlu +++ b/src/kernels/mlu/src/highSoftmax_device.mlu @@ -45,7 +45,7 @@ __mlu_device__ void softmaxKernelAxis_m(float* destination, float* source, int f __bang_add(tmpSum, tmpSum, src, maxNum);//sum += exp(x - M) __memcpy(tmpOldMax, tmpNewMax, maxNum * sizeof(float), NRAM2NRAM);//oldM = newM } - __bang_active_recip(tmpSum, tmpSum, maxNum);//计算1/sum + __bang_active_reciphp(tmpSum, tmpSum, maxNum);//计算1/sum //开始指数变换并且写回GDRAM __bang_mul(src, src, tmpSum, maxNum);//上面循环结束src存储的数据可以利用 __memcpy(destination + (dimsize - 1) * stride + frontIdx + j * maxNum, src, maxNum * sizeof(float), NRAM2GDRAM); @@ -75,7 +75,7 @@ __mlu_device__ void softmaxKernelAxis_m(float* destination, float* source, int f __memcpy(tmpOldMax, tmpNewMax, maxNum * sizeof(float), NRAM2NRAM);//oldM = newM } //------------------- - __bang_active_recip(tmpSum, tmpSum, maxNum);//计算1/sum + __bang_active_reciphp(tmpSum, tmpSum, maxNum);//计算1/sum //开始指数变换并且写回GDRAM __bang_mul(src, src, tmpSum, maxNum);//上面循环结束src存储的数据可以利用 __memcpy(destination + (dimsize - 1) * stride + frontIdx + repeat * maxNum, src, remain * sizeof(float), NRAM2GDRAM); @@ -157,7 +157,7 @@ __mlu_device__ void softmaxKernelAxis_m(float* destination, float* source, int f //此时tmpNewMax存储的是对应于固定frontIdx，behindsize对应数据的最大值，而tmpSum存储的就是对应数值和 //__bang_printf("tmpOldMax[0]:%.2f,tmpSum[0]:%.2f\n", tmpNewMax[2],tmpSum[2]); - __bang_active_recip(tmpSum, tmpSum, strideS); + __bang_active_reciphp(tmpSum, tmpSum, strideS); //__bang_printf("tmpOldMax[0]:%.2f,tmpSum[0]:%.2f\n", tmpNewMax[2],tmpSum[2]); if(remain){ for(int m = 0; m < remain; m++){ @@ -225,7 +225,7 @@ __mlu_device__ void softmaxKernelAxis_m(float* destination, float* source, int f __bang_add(tmpSum, tmpSum, tmp, strideS);//sum += exp(x - M) __memcpy(tmpOldMax, tmpNewMax, stride * sizeof(float), NRAM2NRAM);//oldM = newM } - __bang_active_recip(tmpSum, tmpSum, strideS); + __bang_active_reciphp(tmpSum, tmpSum, strideS); __bang_mul(tmp, tmp, tmpSum, strideS);//上面循环结束tmp存储的数据可以利用 //__memcpy(destination + tid + m * behindsize + (dimsize - 1) * stride, tmp, stride * sizeof(float), NRAM2GDRAM); __memcpy(src + m * behindsize + (dimsize - 1) * stride, tmp, stride * sizeof(float), NRAM2NRAM); @@ -262,7 +262,7 @@ __mlu_device__ void softmaxKernelAxis_m(float* destination, float* source, int f __memcpy(tmpOldMax, tmpNewMax, stride * sizeof(float), NRAM2NRAM);//oldM = newM } //__bang_printf("max:%.2f,%.2f, sum:%.2f,sum:%.2f\n", tmpNewMax[0], tmpNewMax[1], tmpSum[0], tmpSum[0]); - __bang_active_recip(tmpSum, tmpSum, strideS); + __bang_active_reciphp(tmpSum, tmpSum, strideS); __bang_mul(tmp, tmp, tmpSum, strideS);//上面循环结束tmp存储的数据可以利用 //__memcpy(destination + tid + m * behindsize + (dimsize - 1) * stride, tmp, stride * sizeof(float), NRAM2GDRAM); __memcpy(src + m * behindsize + (dimsize - 1) * stride, tmp, stride * sizeof(float), NRAM2NRAM); @@ -473,7 +473,7 @@ __mlu_device__ void softmaxKernelAxis_s(float* destination, float* source, int o __bang_add(tmpSum, tmpSum, src, maxNum);//sum += exp(x - M) __memcpy(tmpOldMax, tmpNewMax, maxNum * sizeof(float), NRAM2NRAM);//oldM = newM } - __bang_active_recip(tmpSum, tmpSum, maxNum);//计算1/sum + __bang_active_reciphp(tmpSum, tmpSum, maxNum);//计算1/sum //开始指数变换并且写回GDRAM __bang_mul(src, src, tmpSum, maxNum);//上面循环结束src存储的数据可以利用 __memcpy(destination + (dimsize - 1) * stride + indStart + j * maxNum, src, maxNum * sizeof(float), NRAM2GDRAM); @@ -505,7 +505,7 @@ __mlu_device__ void softmaxKernelAxis_s(float* destination, float* source, int o __memcpy(tmpOldMax, tmpNewMax, maxNum * sizeof(float), NRAM2NRAM);//oldM = newM } - __bang_active_recip(tmpSum, tmpSum, maxNum);//计算1/sum + __bang_active_reciphp(tmpSum, tmpSum, maxNum);//计算1/sum //开始指数变换并且写回GDRAM __bang_mul(src, src, tmpSum, maxNum);//上面循环结束src存储的数据可以利用 __memcpy(destination + (dimsize - 1) * stride + indStart + repeat * maxNum, src, remainNram * sizeof(float), NRAM2GDRAM); diff --git a/test/kernels/bang/test_bang_softmax.cc b/test/kernels/bang/test_bang_softmax.cc index 0ce65776..83fcd0e8 100644 --- a/test/kernels/bang/test_bang_softmax.cc +++ b/test/kernels/bang/test_bang_softmax.cc @@ -6,7 +6,7 @@ #include "test.h" #include namespace infini { - +double eps = 3e-3; TEST(cuDNN_Softmax, run_axis1) { // Runtime Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance(); @@ -28,7 +28,8 @@ TEST(cuDNN_Softmax, run_axis1) { // Check EXPECT_TRUE(outputGpu2Cpu->equalData( vector{0.032058604, 0.08714432, 0.23688284, 0.6439143, - 0.032058604, 0.08714432, 0.23688284, 0.6439143})); + 0.032058604, 0.08714432, 0.23688284, 0.6439143}, + eps)); } TEST(cuDNN_Softmax, run_axis0) { @@ -50,8 +51,8 @@ TEST(cuDNN_Softmax, run_axis0) { auto outputGpu = gpuOp->getOutput(); auto outputGpu2Cpu = outputGpu->clone(cpuRuntime); // Check - EXPECT_TRUE( - outputGpu2Cpu->equalData(vector{0., 0., 0., 0., 1, 1, 1, 1})); + EXPECT_TRUE(outputGpu2Cpu->equalData( + vector{0., 0., 0., 0., 1, 1, 1, 1}, eps)); } TEST(cuDNN_Softmax2, run_axis1) { @@ -73,10 +74,12 @@ TEST(cuDNN_Softmax2, run_axis1) { auto outputGpu = gpuOp->getOutput(); auto outputGpu2Cpu = outputGpu->clone(cpuRuntime); // Check - EXPECT_TRUE(outputGpu2Cpu->equalData(vector{ - 0.0179862, 0.0179862, 0.0179862, 0.0179862, 0.9820138, 0.9820138, - 0.9820138, 0.9820138, 0.0179862, 0.0179862, 0.0179862, 0.0179862, - 0.9820138, 0.9820138, 0.9820138, 0.9820138})); + EXPECT_TRUE(outputGpu2Cpu->equalData( + vector{0.0179862, 0.0179862, 0.0179862, 0.0179862, 0.9820138, + 0.9820138, 0.9820138, 0.9820138, 0.0179862, 0.0179862, + 0.0179862, 0.0179862, 0.9820138, 0.9820138, 0.9820138, + 0.9820138}, + eps)); } TEST(cuDNN_Softmax2, run_axis2) { @@ -98,10 +101,12 @@ TEST(cuDNN_Softmax2, run_axis2) { auto outputGpu = gpuOp->getOutput(); auto outputGpu2Cpu = outputGpu->clone(cpuRuntime); // Check - EXPECT_TRUE(outputGpu2Cpu->equalData(vector{ - 0.1192029, 0.1192029, 0.8807971, 0.8807971, 0.1192029, 0.1192029, - 0.8807971, 0.8807971, 0.1192029, 0.1192029, 0.8807971, 0.8807971, - 0.1192029, 0.1192029, 0.8807971, 0.8807971})); + EXPECT_TRUE(outputGpu2Cpu->equalData( + vector{0.1192029, 0.1192029, 0.8807971, 0.8807971, 0.1192029, + 0.1192029, 0.8807971, 0.8807971, 0.1192029, 0.1192029, + 0.8807971, 0.8807971, 0.1192029, 0.1192029, 0.8807971, + 0.8807971}, + eps)); } TEST(cuDNN_Softmax2, run_axis3) { @@ -123,9 +128,11 @@ TEST(cuDNN_Softmax2, run_axis3) { auto outputGpu = gpuOp->getOutput(); auto outputGpu2Cpu = outputGpu->clone(cpuRuntime); // Check - EXPECT_TRUE(outputGpu2Cpu->equalData(vector{ - 0.2689414, 0.7310586, 0.2689414, 0.7310586, 0.2689414, 0.7310586, - 0.2689414, 0.7310586, 0.2689414, 0.7310586, 0.2689414, 0.7310586, - 0.2689414, 0.7310586, 0.2689414, 0.7310586})); + EXPECT_TRUE(outputGpu2Cpu->equalData( + vector{0.2689414, 0.7310586, 0.2689414, 0.7310586, 0.2689414, + 0.7310586, 0.2689414, 0.7310586, 0.2689414, 0.7310586, + 0.2689414, 0.7310586, 0.2689414, 0.7310586, 0.2689414, + 0.7310586}, + eps)); } } // namespace infini