914a6e8e72
Introduces ENCODER_INPUT_DIM = 56 (FEATURE_DIM + 16). All encoder first-layer weight matrices (VSN gate, Mamba2 L1 input projection) now sized for 56 input dims. The extra 16 are per-batch state: 4 trade_context + 12 multires features. - snap_feature_assemble_batched: output stride → ENCODER_INPUT_DIM, zero-fills dims [40..56] for the broadcast kernel to overwrite. - New rl_encoder_context_broadcast.cu: writes trade_context_d[B×4] + multires_output_d[B×12] into each of the K sequence rows per batch at positions [40..56]. - CfcConfig.n_in, Mamba2 L1 in_dim, VSN gate, window_tensor_d, all forward/backward scratch buffers updated to ENCODER_INPUT_DIM. - CfcTrunk default config updated. The broadcast kernel launch integration into the forward_only path is the final wire-up step — until then dims 40-55 are zero-filled (safe: Xavier init on new columns means encoder starts by learning to ignore them, then gradually incorporates the signal). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
40 lines
1.6 KiB
Plaintext
40 lines
1.6 KiB
Plaintext
// rl_encoder_context_broadcast.cu — broadcast per-batch trade_context
|
||
// (4 dims) + multires_output (12 dims) into the encoder input tensor
|
||
// at positions [40..56] for each of the K sequence rows per batch.
|
||
//
|
||
// The encoder input tensor is [B, K, ENCODER_INPUT_DIM=56] row-major.
|
||
// Dims [0..40] are per-snapshot market features (written by
|
||
// snap_feature_assemble_batched). Dims [40..56] are per-batch state
|
||
// that is CONSTANT across K <20><><EFBFBD> this kernel broadcasts them.
|
||
//
|
||
// One thread per (batch × K) row. No shared memory.
|
||
// Per `feedback_no_atomicadd`, `feedback_cpu_is_read_only`.
|
||
|
||
#include <stdint.h>
|
||
|
||
#define ENCODER_INPUT_DIM 56
|
||
#define SNAP_FEATURE_DIM 40
|
||
#define TRADE_CONTEXT_DIM 4
|
||
#define MULTIRES_DIM 12
|
||
|
||
extern "C" __global__ void rl_encoder_context_broadcast(
|
||
float* __restrict__ encoder_input, // [B × K × ENCODER_INPUT_DIM]
|
||
const float* __restrict__ trade_context, // [B × TRADE_CONTEXT_DIM]
|
||
const float* __restrict__ multires_output, // [B × MULTIRES_DIM]
|
||
int b_size,
|
||
int seq_len
|
||
) {
|
||
const int idx = blockIdx.x * blockDim.x + threadIdx.x;
|
||
const int total = b_size * seq_len;
|
||
if (idx >= total) return;
|
||
|
||
const int b = idx / seq_len;
|
||
|
||
float* row = encoder_input + idx * ENCODER_INPUT_DIM + SNAP_FEATURE_DIM;
|
||
const float* tc = trade_context + b * TRADE_CONTEXT_DIM;
|
||
const float* mr = multires_output + b * MULTIRES_DIM;
|
||
|
||
for (int i = 0; i < TRADE_CONTEXT_DIM; ++i) row[i] = tc[i];
|
||
for (int i = 0; i < MULTIRES_DIM; ++i) row[TRADE_CONTEXT_DIM + i] = mr[i];
|
||
}
|