From 0afe61f88ce3d2f445c5f01ae5567cb1b0b7f303 Mon Sep 17 00:00:00 2001 From: Eric Kunze Date: Wed, 14 Feb 2024 16:33:31 -0800 Subject: Modify convolution operators to improve bias handling Accumulator size moves to an enumerated attribute, out_t for floating-point changes to be the size of the input. Bias for floating-point also becomes the bit width of the input type. Signed-off-by: Eric Kunze Change-Id: I7369417adbb1106ce34a1978e7f511a30272c318 --- chapters/introduction.adoc | 4 +- pseudocode/operators/CONV2D.tosac | 17 ++--- pseudocode/operators/CONV3D.tosac | 5 +- pseudocode/operators/DEPTHWISE_CONV2D.tosac | 5 +- pseudocode/operators/TRANSPOSE_CONV2D.tosac | 34 +++++----- tosa.xml | 96 +++++++++++++++++------------ tosa.xsd | 2 +- 7 files changed, 94 insertions(+), 69 deletions(-) diff --git a/chapters/introduction.adoc b/chapters/introduction.adoc index c34bf7b..c6764d3 100644 --- a/chapters/introduction.adoc +++ b/chapters/introduction.adoc @@ -397,7 +397,7 @@ if (!local_bound) { output_bnd = operation_fp64(input_abs, weight_abs, bias_abs); size_t T = tensor_size(output_shape) // number dot product results -size_t ksb = (max_value(bias_abs) > 0) ? (KS + 1) : KS; // kernel size and bias +size ksb = ceil(KS / exp2(normal_frac() - normal_frac())) + ((max_value(bias_abs) > 0) ? 1 : 0); fp64_t out_err_sum = 0.0; fp64_t out_err_sumsq = 0.0; for_each(index in output_shape) { @@ -412,7 +412,7 @@ for_each(index in output_shape) { REQUIRE(out_ref == 0.0 && out_imp == 0.0); out_err = 0.0; } else { // 0.0 < out_bnd < infinity - fp64_t out_err_bnd = max(out_bnd * exp2(-1-normal_frac()), normal_min()); + fp64_t out_err_bnd = max(out_bnd * exp2(-1-normal_frac()), normal_min()); out_err = (static_cast(out_imp) - out_ref) / out_err_bnd; REQUIRE(abs(out_err) <= ksb); } diff --git a/pseudocode/operators/CONV2D.tosac b/pseudocode/operators/CONV2D.tosac index fe61747..0ae0e81 100644 --- a/pseudocode/operators/CONV2D.tosac +++ b/pseudocode/operators/CONV2D.tosac @@ -17,24 +17,25 @@ ERROR_IF(OW != idiv_check(IW - 1 + pad_left + pad_right - (KW - 1) * dilation_x, ERROR_IF(BC != OC && BC != 1); for_each(0 <= n < N, 0 <= oy < OH, 0 <= ox < OW, 0 <= oc < OC) { - out_t acc = 0; + acc_t acc = 0; index_t iy = oy * stride_y - pad_top; index_t ix = ox * stride_x - pad_left; for_each(0 <= ky < KH, 0 <= kx < KW, 0 <= ic < IC) { index_t y = iy + ky * dilation_y; index_t x = ix + kx * dilation_x; if (0 <= y < IH && 0 <= x < IW) { - out_t value = static_cast(tensor_read(input, + acc_t value = static_cast(tensor_read(input, [N,IH,IW,IC], [n,y,x,ic])); - out_t weight = static_cast(tensor_read(weight, + acc_t weight = static_cast(tensor_read(weight, [OC,KH,KW,IC], [oc,ky,kx,ic])); - value = apply_sub_s(value, static_cast(input_zp)); - weight = apply_sub_s(weight, static_cast(weight_zp)); - acc = apply_add_s(acc, apply_mul_s(value, weight)); + value = apply_sub_s(value, static_cast(input_zp)); + weight = apply_sub_s(weight, static_cast(weight_zp)); + acc = apply_add_s(acc, apply_mul_s(value, weight)); } } - acc = apply_add_s(acc, bias[(BC == 1) ? 0 : oc]); - tensor_write(output, [N,OH,OW,OC], [n,oy,ox,oc], acc); + out_t out = static_cast(acc); + out = apply_add_s(out, bias[(BC == 1) ? 0 : oc]); + tensor_write(output, [N,OH,OW,OC], [n,oy,ox,oc], out); } diff --git a/pseudocode/operators/CONV3D.tosac b/pseudocode/operators/CONV3D.tosac index 7568564..e53b7eb 100644 --- a/pseudocode/operators/CONV3D.tosac +++ b/pseudocode/operators/CONV3D.tosac @@ -38,6 +38,7 @@ for_each(0 <= n < N, 0 <= od < OD, 0 <= oy < OH, 0 <= ox < OW, 0 <= oc < OC) { acc = apply_add_s(acc, apply_mul_s(value, weight)); } } - acc = apply_add_s(acc, bias[(BC == 1) ? 0 : oc]); - tensor_write(output, [N,OD,OH,OW,OC], [n,od,oy,ox,oc], acc); + out_t out = static_cast(acc); + out = apply_add_s(out, bias[(BC == 1) ? 0 : oc]); + tensor_write(output, [N,OD,OH,OW,OC], [n,od,oy,ox,oc], out); } diff --git a/pseudocode/operators/DEPTHWISE_CONV2D.tosac b/pseudocode/operators/DEPTHWISE_CONV2D.tosac index a473375..419d2eb 100644 --- a/pseudocode/operators/DEPTHWISE_CONV2D.tosac +++ b/pseudocode/operators/DEPTHWISE_CONV2D.tosac @@ -35,6 +35,7 @@ for_each(0 <= n < N, 0 <= oy < OH, 0 <= ox < OW, 0 <= c < C, 0 <= m < M) { acc = apply_add_s(acc, apply_mul_s(value, weight)); } } - acc = apply_add_s(acc, bias[(BC == 1) ? 0 : (c * M) + m]); - tensor_write(output, [N,OH,OW,C * M], [n,oy,ox,c * M + m], acc); + out_t out = static_cast(acc); + out = apply_add_s(out, bias[(BC == 1) ? 0 : (c * M) + m]); + tensor_write(output, [N,OH,OW,C * M], [n,oy,ox,c * M + m], out); } diff --git a/pseudocode/operators/TRANSPOSE_CONV2D.tosac b/pseudocode/operators/TRANSPOSE_CONV2D.tosac index ab61348..6713b30 100644 --- a/pseudocode/operators/TRANSPOSE_CONV2D.tosac +++ b/pseudocode/operators/TRANSPOSE_CONV2D.tosac @@ -16,20 +16,24 @@ ERROR_IF(OH != (IH - 1) * stride_y + out_pad_top + out_pad_bottom + KH); ERROR_IF(OW != (IW - 1) * stride_x + out_pad_left + out_pad_right + KW); ERROR_IF(BC != OC && BC != 1); -for_each(index in [N, OH, OW, OC]) { - tensor_write(output, [N,OH,OW,OC], index, bias[(BC == 1) ? 0 : index[3]]); -} -for_each(0 <= n < N, 0 <= iy < IH, 0 <= ix < IW, 0 <= oc < OC, - 0 <= ic < IC, 0 <= ky < KH, 0 <= kx < KW) { - index_t oy = iy * stride_y + out_pad_top + ky; - index_t ox = ix * stride_x + out_pad_left + kx; - if (oy >= 0 && oy < OH && ox >= 0 && ox < OW) { - out_t acc = static_cast(tensor_read(output, [N,OH,OW,OC], [n,oy,ox,oc])); - out_t value = static_cast(tensor_read(input, [N,IH,IW,IC], [n,iy,ix,ic])); - out_t weight = static_cast(tensor_read(weight, [OC,KH,KW,IC], [oc,ky,kx,ic])); - value = apply_sub_s(value, static_cast(input_zp)); - weight = apply_sub_s(weight, static_cast(weight_zp)); - acc = apply_add_s(acc, apply_mul_s(value, weight)); - tensor_write(output, [N,OH,OW,OC], [n,oy,ox,oc], acc); +for_each(0 <= n < N, 0 <= iy < IH, 0 <= ix < IW, 0 <= dy < stride_y, 0 <= dx < stride_x, 0 <= oc < OC) { + acc_t acc = 0; + index_t oy = iy * stride_y + dy + out_pad_top; + index_t ox = ix * stride_x + dx + out_pad_left; + + for_each(0 <= sy * stride_y < KY - dy, 0 <= sx * stride_x < KX - dx, 0 <= ic < IC) { + index_t y = iy - sy; + index_t x = ix - sx; + index_t ky = dy + sy * stride_y; + index_t kx = dx + sx * stride_x; + acc_t value = static_cast(tensor_read(input, [N,IH,IW,IC], [n,y,x,ic])); + acc_t weight_value = static_cast(tensor_read(weight, [OH,KH,KW,IC], [oc,ky,kx,ic])); + value = apply_sub_s(value, static_cast(input_zp)); + weight_value = apply_sub_s(weight_value, static_cast(weight_zp)); + acc = apply_add_s(acc, apply_mul_s(value, weight_value)); } + + out_t out = static_cast(acc); + out = apply_add_s(out, bias[(BC == 1) ? 0 : oc]); + tensor_write(output, [N,OH,OW,OC], [n,oy,ox,oc], out); } diff --git a/tosa.xml b/tosa.xml index 17c82f9..675eeb6 100644 --- a/tosa.xml +++ b/tosa.xml @@ -92,8 +92,8 @@ - - Enumerated type, must be one of INT32, FP16, FP32, as defined in the Supported Data Types table for this operation + + Enumerated type, must be one of INT32, INT48, FP16, FP32, as defined in the Supported Data Types table for this operation @@ -174,6 +174,10 @@ [dilation_y, dilation_x] + + Enumerated type, must be one of INT32, INT48, FP16, FP32, as defined in the Supported Data Types table for this operation + + Input tensor zero point. Must be zero for non-int8 types. @@ -199,32 +203,33 @@ + - + - + - + - + - + - + - + - + - + @@ -268,6 +273,10 @@ [dilation_d, dilation_y, dilation_x] + + Enumerated type, must be one of INT32, INT48, FP16, FP32, as defined in the Supported Data Types table for this operation + + Input tensor zero point. Must be zero for non-int8 types. @@ -294,31 +303,31 @@ - + - + - + - + - + - + - + - + - + @@ -358,6 +367,10 @@ [dilation_y, dilation_x] + + Enumerated type, must be one of INT32, INT48, FP16, FP32, as defined in the Supported Data Types table for this operation + + Input tensor zero point. Must be zero for non-int8 types. @@ -384,31 +397,31 @@ - + - + - + - + - + - + - + - + - + @@ -684,6 +697,10 @@ + + Enumerated type, must be one of INT32, INT48, FP16, FP32, as defined in the Supported Data Types table for this operation + + Input tensor zero point. Must be zero for non-int8 types. @@ -704,37 +721,37 @@ Output tensor - + - + - + - + - + - + - + - + - + @@ -3299,10 +3316,11 @@ used. - + + diff --git a/tosa.xsd b/tosa.xsd index b39a2f4..54610d5 100644 --- a/tosa.xsd +++ b/tosa.xsd @@ -57,7 +57,7 @@ - + -- cgit v1.2.1