diff --git a/hpvm/projects/hpvm-tensor-rt/tensor_runtime/src/approx_simulation.cu b/hpvm/projects/hpvm-tensor-rt/tensor_runtime/src/approx_simulation.cu
index a472fcaa36484950de98f858a74f185900ab80b7..a9613b347ecbedba1e387a73f47756d478f00aed 100644
--- a/hpvm/projects/hpvm-tensor-rt/tensor_runtime/src/approx_simulation.cu
+++ b/hpvm/projects/hpvm-tensor-rt/tensor_runtime/src/approx_simulation.cu
@@ -532,643 +532,6 @@ void *tensorConvSampSim2(void *input_ptr, void *filter_ptr, int vertical_pad,
return output;
}
-/************ NOTE: API for ApproxHPVM Wrapper runtime *******/
-void *PROMISE_Conv(void *input, float i_min, float i_max, void *filter,
- float w_min, float w_max, void *bias, float b_min,
- float b_max, int conv_pad_h, int conv_pad_w,
- int conv_stride_h, int conv_stride_w, int pool_id,
- int pool_size, int pool_stride,
- int activation_id, // Relu, Tanh, ClipRelu
- float out_min, float out_max, int swing) {
-
- Tensor *input_t = (Tensor *)input;
- Tensor *filter_t = (Tensor *)filter;
- Tensor *bias_t = (Tensor *)bias;
-
- int orig_type = input_t->cur_type;
-
- DEBUG("FP32 conversions \n");
-
- convertToFP32(input_t);
-
- convertToFP32(filter_t);
- convertToFP32(bias_t);
-
- DEBUG("DONE FP32 conversions \n");
-
- if (swing < 8) {
- input = quantizeTensorPromise(input, i_min, i_max);
- filter = quantizeTensorPromise(filter, w_min, w_max);
- if (bias != NULL)
- bias = quantizeTensorPromise(bias, b_min, b_max);
- // aRead error
-
- input = addPromiseError(input, swing);
- }
-
- void *conv_out;
- conv_out = tensorConvolution(input, filter, conv_pad_h, conv_pad_w,
- conv_stride_h, conv_stride_w, 1, 0);
-
- void *conv_add;
- if (bias != NULL) {
- conv_add = tensorAdd(conv_out, bias);
- } else {
- conv_add = conv_out;
- }
-
- void *pool_out;
- // NOTE: Skip pooling on negative pool sizes
- if (pool_size > 0) {
- // FIXME: Currently only using MaxPooling
- //-- pool_out = tensorPooling(conv_add, 0, pool_size, pool_size, 0, 0,
- // pool_size, pool_size);
- pool_out = tensorPooling(conv_add, 0, pool_size, pool_size, 0, 0,
- pool_stride, pool_stride);
- } else {
- pool_out = conv_add;
- }
-
- void *activation_out;
- switch (activation_id) {
- case -1:
- activation_out = pool_out;
- INFO("NO Activation Function \n");
- break;
- case 0:
- activation_out = tensorTanh(pool_out);
- break;
- case 1:
- activation_out = tensorRelu(pool_out);
- break;
- case 2:
- activation_out = tensorRelu2(pool_out, out_min, out_max);
- break;
- default:
- ERROR("Activation id %d NOT supported \n", activation_out);
- break;
- }
-
- if (swing < 8 && activation_id != -1) {
- activation_out = quantizeTensorPromise(activation_out, out_min, out_max);
- }
-
- // NOTE: Convert back to FP16 if original type
- if (orig_type == half_type) {
- convertToFP16((Tensor *)activation_out);
- }
-
- return activation_out;
-}
-
-void *PROMISE_FC(void *input, float i_min, float i_max, void *weights,
- float w_min, float w_max, void *bias, float b_min, float b_max,
- int activation_id, float out_min, float out_max, int swing) {
-
- Tensor *input_t = (Tensor *)input;
- Tensor *weights_t = (Tensor *)weights;
- Tensor *bias_t = (Tensor *)bias;
-
- int orig_type = input_t->cur_type;
-
- convertToFP32(input_t);
- convertToFP32(weights_t);
- convertToFP32(bias_t);
-
- if (swing < 8) {
- input = quantizeTensorPromise(input, i_min, i_max);
- weights = quantizeTensorPromise(weights, w_min, w_max);
- if (bias != NULL)
- bias = quantizeTensorPromise(bias, b_min, b_max);
-
- // NOTE: Modelling aRead error in PROMISE
- input = addPromiseError(input, swing);
- }
-
- void *gemm_out;
- gemm_out = tensorGemmGPU(input, weights);
-
- void *gemmbias_out;
- if (bias != NULL) {
- gemmbias_out = tensorAdd(gemm_out, bias);
- } else {
- gemmbias_out = gemm_out;
- }
-
- void *activation_out;
- switch (activation_id) {
-
- case -1:
- activation_out = gemmbias_out;
- INFO("No Activation Function \n");
- break;
- case 0:
- activation_out = tensorTanh(gemmbias_out);
- break;
- case 1:
- activation_out = tensorRelu(gemmbias_out);
- break;
- case 2:
- activation_out = tensorRelu2(gemmbias_out, out_min, out_max);
- break;
- default:
- ERROR("Activation id %d NOT supported \n", activation_out);
- break;
- }
-
- if (swing < 8 && activation_id != -1) {
- activation_out = quantizeTensorPromise(activation_out, out_min, out_max);
- }
-
- // NOTE: Convert back to FP16 if original type
- if (orig_type == half_type) {
- convertToFP16((Tensor *)activation_out);
- }
-
- return activation_out;
-}
-
-// NOTE: Enabling the macro below is used for testing against the old PROMISE
-// wrapper
-//#define OLD_MODEL
-
-#ifndef OLD_MODEL
-
-bool isPromiseLayer(int swing) {
-
- if (swing < 8)
- return true;
- else
- return false;
-}
-
-bool isGPULayer(int swing) {
-
- if (swing > 10) // PROMISE layers are 1-7
- return true;
- else
- return false;
-}
-
-bool isFullPrecision(int swing) {
-
- if (swing == 11)
- return true;
- else
- return false;
-}
-
-bool isHalfPrecision(int swing) {
-
- if (swing == 12)
- return true;
- else
- return false;
-}
-
-bool isPerforation(int swing) {
-
- if (swing >= 100 && swing <= 200)
- return true;
- else
- return false;
-}
-
-bool isSampling(int swing) {
-
- if (swing >= 200 && swing <= 300)
- return true;
- else
- return false;
-}
-
-bool isReductionSampling(int swing) {
-
- if (swing >= 41 && swing <= 49)
- return true;
- else
- return false;
-}
-
-int getSwing(int swing) {
-
-#ifdef PROMISE_TUNER_ENABLED
-
- // NOTE: Skip reading file-based error levels for ApproxHPVM wrapper runtime
- if (!approxhpvm_runtime_mode) {
-
- if (op_counter >= total_ops) {
- ERROR("No accuracy flag found \n");
- }
-
- swing = op_accuracies[op_counter];
- op_counter++;
- }
-
-#endif
-
- DEBUG("---- swing_value = %d \n", swing);
-
- return swing;
-}
-
-// bool FP16_tuning = false;
-
-/***** API for Autotuner Use - Not the ApproxHPVM Wrapper API */
-
-void initializeAutotuner() {
-
- DEBUG("initializing tuner .... \n");
-
- sampParamSet = new SampParamSet;
- perfParamSet = new PerfParamSet;
-}
-
-void *Autotuner_SampConv(void *input, float i_min, float i_max, void *filter,
- float w_min, float w_max, void *bias, float b_min,
- float b_max, int conv_pad_h, int conv_pad_w,
- int conv_stride_h, int conv_stride_w, int pool_id,
- int pool_size,
- int activation_id, // Relu, Tanh, ClipRelu
- float out_min, float out_max, int swing) {
-
- SampParams params = sampParamSet->getSampParams(swing);
-
- DEBUG("params.skip_rate = %d, params.skip_offset = %d \n", params.skip_rate,
- params.skip_offset);
-
- void *conv_out;
-
- if (!FP16_tuning) {
-
- /* conv_out = tensorConvSampSim(input, filter,
- conv_pad_h, conv_pad_w,
- conv_stride_h, conv_stride_w, 1, 1,
- params.skip_rate, params.skip_offset);
- */
-
- if (SIMULATION_MODE) {
- conv_out = tensorConvSampSim2(
- input, filter, conv_pad_h, conv_pad_w, conv_stride_h, conv_stride_w,
- 1, 1, params.skip_rate, params.skip_offset, params.interpolation_id);
- }
-
- else {
- conv_out = tensorConvApprox(input, filter, conv_pad_h, conv_pad_w,
- conv_stride_h, conv_stride_w, 1, 1, 1, 1,
- params.skip_rate, params.skip_offset);
- }
-
- } else {
-
- conv_out = tensorConvApproxHalf2(input, filter, conv_pad_h, conv_pad_w,
- conv_stride_h, conv_stride_w, 1, 1, 1, 1,
- params.skip_rate, params.skip_offset);
- }
-
- return conv_out;
-}
-
-void *Autotuner_PerforatedConv(void *input, float i_min, float i_max,
- void *filter, float w_min, float w_max,
- void *bias, float b_min, float b_max,
- int conv_pad_h, int conv_pad_w,
- int conv_stride_h, int conv_stride_w,
- int pool_id, int pool_size,
- int activation_id, // Relu, Tanh, ClipRelu
- float out_min, float out_max, int swing) {
-
- PerfParams params = perfParamSet->getPerfParams(swing);
-
- DEBUG("params.row = %d, params.col = %d, params.skip_offset = %d \n",
- params.row, params.col, params.skip_offset);
-
- void *conv_out;
-
- if (!FP16_tuning) {
-
- if (SIMULATION_MODE) {
-
- conv_out = tensorConvPerfCuda(input, filter, conv_pad_h, conv_pad_w,
- conv_stride_h, conv_stride_w, 1, 1,
- params.row, params.col, params.skip_offset);
-
- } else {
-
- conv_out = tensorConvApprox(
- input, filter, conv_pad_h, conv_pad_w, conv_stride_h, conv_stride_w,
- 1, 1, params.row, params.col, 1, params.skip_offset);
- }
-
- } else {
-
- conv_out = tensorConvApproxHalf2(
- input, filter, conv_pad_h, conv_pad_w, conv_stride_h, conv_stride_w, 1,
- 1, params.row, params.col, 1, params.skip_offset);
- }
-
- return conv_out;
-}
-
-void *Autotuner_ConvOp(void *input, float i_min, float i_max, void *filter,
- float w_min, float w_max, void *bias, float b_min,
- float b_max, int conv_pad_h, int conv_pad_w,
- int conv_stride_h, int conv_stride_w, int pool_id,
- int pool_size,
- int activation_id, // Relu, Tanh, ClipRelu
- float out_min, float out_max, int swing) {
-
- void *conv_out;
- if (isPerforation(swing)) {
-
- conv_out = Autotuner_PerforatedConv(
- input, i_min, i_max, filter, w_min, w_max, bias, b_min, b_max,
- conv_pad_h, conv_pad_w, conv_stride_h, conv_stride_w, pool_id,
- pool_size, activation_id, out_min, out_max, swing);
-
- }
-
- else if (isSampling(swing)) {
-
- conv_out = Autotuner_SampConv(
- input, i_min, i_max, filter, w_min, w_max, bias, b_min, b_max,
- conv_pad_h, conv_pad_w, conv_stride_h, conv_stride_w, pool_id,
- pool_size, activation_id, out_min, out_max, swing);
- }
-
- else if (isHalfPrecision(swing)) {
-
- if (FP16_tuning) {
-
- conv_out = tensorHalfConvolution(input, filter, conv_pad_h, conv_pad_w,
- conv_stride_h, conv_stride_w, 1, 0);
- } else {
- conv_out = tensorConvolution(input, filter, conv_pad_h, conv_pad_w,
- conv_stride_h, conv_stride_w, 1, 0);
- }
-
- }
-
- else if (isFullPrecision(swing)) {
- conv_out = tensorConvolution(input, filter, conv_pad_h, conv_pad_w,
- conv_stride_h, conv_stride_w, 1, 0);
- }
-
- return conv_out;
-}
-
-void *Autotuner_Add(void *input, void *bias, int swing) {
-
- void *conv_add;
- if (bias != NULL) {
-
- if (isFullPrecision(swing) || !(FP16_tuning)) {
- conv_add = tensorAdd(input, bias);
- } else {
- conv_add = tensorHalfAdd(input, bias);
- }
- } else {
- conv_add = input;
- }
-
- return conv_add;
-}
-
-void *Autotuner_Pooling(void *input, int pool_size, int pool_stride,
- int swing) {
-
- void *pool_out;
-
- if (pool_size > 0) {
-
- // FIXME: Currently only using MaxPooling
- if (isFullPrecision(swing) || !(FP16_tuning)) {
- pool_out = tensorPooling(input, 0, pool_size, pool_size, 0, 0,
- pool_stride, pool_stride);
-
- }
-
- else {
- pool_out = tensorHalfPooling(input, 0, pool_size, pool_size, 0, 0,
- pool_stride, pool_stride);
- }
-
- } else {
- pool_out = input;
- }
-
- return pool_out;
-}
-
-void *Autotuner_Activation(void *input, int activation_id, int out_min,
- int out_max, int swing) {
-
- void *activation_out;
-
- if (isFullPrecision(swing) || (!FP16_tuning)) {
-
- switch (activation_id) {
- case -1:
- activation_out = input;
- INFO("NO Activation Function \n");
- break;
- case 0:
- activation_out = tensorTanh(input);
- break;
- case 1:
- activation_out = tensorRelu(input);
- break;
- case 2:
- activation_out = tensorRelu2(input, out_min, out_max);
- break;
- default:
- ERROR("Activation id %d NOT supported \n", activation_out);
- break;
- }
- }
-
- else {
-
- switch (activation_id) {
- case -1:
- activation_out = input;
- INFO("NO Activation Function \n");
- break;
- case 0:
- activation_out = tensorHalfTanh(input);
- break;
- case 1:
- activation_out = tensorHalfRelu(input);
- break;
- case 2:
- activation_out = tensorHalfRelu2(input, out_min, out_max);
- break;
- default:
- ERROR("Activation id %d NOT supported \n", activation_out);
- break;
- }
- }
-
- return activation_out;
-}
-
-void *Autotuner_GPU_ConvLayer(void *input, float i_min, float i_max,
- void *filter, float w_min, float w_max,
- void *bias, float b_min, float b_max,
- int conv_pad_h, int conv_pad_w, int conv_stride_h,
- int conv_stride_w, int pool_id, int pool_size,
- int pool_stride,
- int activation_id, // Relu, Tanh, ClipRelu
- float out_min, float out_max, int swing) {
-
- void *conv_out = Autotuner_ConvOp(
- input, i_min, i_max, filter, w_min, w_max, bias, b_min, b_max, conv_pad_h,
- conv_pad_w, conv_stride_h, conv_stride_w, pool_id, pool_size,
- activation_id, out_min, out_max, swing);
-
- void *conv_add = Autotuner_Add(conv_out, bias, swing);
-
- void *pool_out = Autotuner_Pooling(conv_add, pool_size, pool_stride, swing);
-
- void *activation_out =
- Autotuner_Activation(pool_out, activation_id, out_min, out_max, swing);
-
- return activation_out;
-}
-
-/**** Top-level API for Handling Convolution Layers
-
- The granularity of handling is at a layer-level - not tensor-op level
-
-***/
-
-void *Autotuner_ConvLayer(void *input, float i_min, float i_max, void *filter,
- float w_min, float w_max, void *bias, float b_min,
- float b_max, int conv_pad_h, int conv_pad_w,
- int conv_stride_h, int conv_stride_w, int pool_id,
- int pool_size, int pool_stride,
- int activation_id, // Relu, Tanh, ClipRelu
- float out_min, float out_max, int swing) {
-
- if (FP16_tuning) {
- if (ONLINE_PROFILING) {
- ERROR("Online Profiling cannot be enabled with PROMISE Simulation \n");
- }
- }
-
- swing = getSwing(swing);
-
- if (isPromiseLayer(swing)) {
-
- return PROMISE_Conv(input, i_min, i_max, filter, w_min, w_max, bias, b_min,
- b_max, conv_pad_h, conv_pad_w, conv_stride_h,
- conv_stride_w, pool_id, pool_size, pool_stride,
- activation_id, out_min, out_max, swing);
- }
-
- assert(isGPULayer(swing));
-
- return Autotuner_GPU_ConvLayer(
- input, i_min, i_max, filter, w_min, w_max, bias, b_min, b_max, conv_pad_h,
- conv_pad_w, conv_stride_h, conv_stride_w, pool_id, pool_size, pool_stride,
- activation_id, out_min, out_max, swing);
-}
-
-/**** Top-level API Unchanged for backwards compatibility ***/
-
-void *ConvLayer_PROMISE(void *input, float i_min, float i_max, void *filter,
- float w_min, float w_max, void *bias, float b_min,
- float b_max, int conv_pad_h, int conv_pad_w,
- int conv_stride_h, int conv_stride_w, int pool_id,
- int pool_size,
- int activation_id, // Relu, Tanh, ClipRelu
- float out_min, float out_max, int swing) {
-
- return Autotuner_ConvLayer(
- input, i_min, i_max, filter, w_min, w_max, bias, b_min, b_max, conv_pad_h,
- conv_pad_w, conv_stride_h, conv_stride_w, pool_id, pool_size,
- pool_size, // FIXIT: Assumption pool_size == pool_strides
- activation_id, out_min, out_max, swing);
-}
-
-void *ConvLayer_PROMISE2(void *input, float i_min, float i_max, void *filter,
- float w_min, float w_max, void *bias, float b_min,
- float b_max, int conv_pad_h, int conv_pad_w,
- int conv_stride_h, int conv_stride_w, int pool_id,
- int pool_size, int pool_stride,
- int activation_id, // Relu, Tanh, ClipRelu
- float out_min, float out_max, int swing) {
-
- return Autotuner_ConvLayer(
- input, i_min, i_max, filter, w_min, w_max, bias, b_min, b_max, conv_pad_h,
- conv_pad_w, conv_stride_h, conv_stride_w, pool_id, pool_size, pool_stride,
- activation_id, out_min, out_max, swing);
-}
-
-void *
-FCLayer_PROMISE(void *input, float i_min, float i_max, void *weights,
- float w_min, float w_max, void *bias, float b_min, float b_max,
- int activation_id, float out_min, float out_max,
- int swing) { // NOTE: min_val, max_val apply to 'ClippedRelu'
-
- swing = getSwing(swing);
-
- if (isPromiseLayer(swing)) {
-
- return PROMISE_FC(input, i_min, i_max, weights, w_min, w_max, bias, b_min,
- b_max, activation_id, out_min, out_max, swing);
- }
-
- assert(isGPULayer(swing));
-
- void *gemm_out;
- if (!isFullPrecision(swing)) {
- gemm_out = tensorHalfGemm(input, weights);
- } else {
- gemm_out = tensorGemmGPU(input, weights);
- }
-
- void *gemmbias_out;
- if (bias != NULL) {
- // Swing 8 corresponds to FP32
- if (isFullPrecision(swing) || (!FP16_tuning)) {
- gemmbias_out = tensorAdd(gemm_out, bias);
- } else {
- gemmbias_out = tensorHalfAdd(gemm_out, bias);
- }
- } else {
- gemmbias_out = gemm_out;
- }
-
- void *activation_out;
- switch (activation_id) {
-
- case -1:
- activation_out = gemmbias_out;
- INFO("No Activation Function \n");
- break;
- case 0:
- activation_out = tensorTanh(gemmbias_out);
- break;
- case 1:
- activation_out = tensorRelu(gemmbias_out);
- break;
- case 2:
- activation_out = tensorRelu2(gemmbias_out, out_min, out_max);
- break;
- default:
- ERROR("Activation id %d NOT supported \n", activation_out);
- break;
- }
-
- return activation_out;
-}
-
-#endif
-
-#ifdef OLD_MODEL
-
-#endif
#endif
diff --git a/hpvm/projects/hpvm-tensor-rt/tensor_runtime/src/init_api.cc b/hpvm/projects/hpvm-tensor-rt/tensor_runtime/src/init_api.cc
index b322ee2be37b60487e15c9109d4230adf1ad84e2..641b342650dce013f0dabd2f1cb4d05c89326798 100644
--- a/hpvm/projects/hpvm-tensor-rt/tensor_runtime/src/init_api.cc
+++ b/hpvm/projects/hpvm-tensor-rt/tensor_runtime/src/init_api.cc
@@ -40,16 +40,6 @@ void llvm_hpvm_initTensorRt(int gpuid) {
DEBUG("CREATED HANDLES %d \n", gpuid);
-#ifdef PROMISE_TUNER_ENABLED
- // readOpenTunerFlags("opentuner_flags");
-
- readOpenTunerFlags("promise_flags");
- initializeAutotuner();
-
- DEBUG("Read PROMISE FLAGS %d \n", gpuid);
-
-#endif
-
runtime_initialized = true;
}