Add compile and operand generate script for flash

kernels/flash_attention/compile_flash.sh

@@ -0,0 +1,46 @@
+#!/bin/bash
+
+archs=("ampere" "virgo")
+
+if [ -z "$TOOLDIR" ]; then
+    echo "error: \$TOOLDIR not set.  Did you run source ci/toolchain_env.sh?"
+    exit 1
+fi
+
+check_exists() {
+    if ! [ -f "$1" ]; then
+        echo "error: looked for file $1 that does not exist."
+        exit 1
+    fi
+}
+
+# generate operands
+echo "generating flash_attn operands for seqlen 1024, headdim 64"
+python3 flash_attn.py 1024 64 64
+mv -v input.a.col.bin input.a.rand.fp32.seqlen1024headdim64.col.bin
+mv -v input.a.row.bin input.a.rand.fp32.seqlen1024headdim64.row.bin
+mv -v input.b.bin input.b.rand.fp32.seqlen1024headdim64.row.bin
+mv -v input.c.bin input.c.rand.fp32.seqlen1024headdim64.row.bin
+ln -sf input.a.rand.fp32.seqlen1024headdim64.row.bin input.a.bin
+ln -sf input.b.rand.fp32.seqlen1024headdim64.row.bin input.b.bin
+ln -sf input.c.rand.fp32.seqlen1024headdim64.row.bin input.c.bin
+
+for arch in "${archs[@]}"; do
+    git checkout ae-$arch
+
+    # re-compile libvortexrt.a
+    # FIXME after restructure
+    pushd ../../lib
+    make
+    popd
+
+    for dim in "${dims[@]}"; do
+        echo "compiling flash_attn kernel for $arch with seqlen 1024, headdim 64"
+
+        # touch source file to force re-building, as the Makefile does not track
+        # binary changes
+        touch kernel.cpp
+
+        make CONFIG=flash.$arch.seqlen1024.headdim64
+    done
+done

kernels/flash_attention/flash_attn.py

@@ -0,0 +1,159 @@
+import sys
+import numpy as np
+
+def parse_mnk():
+    if len(sys.argv) != 4:
+        print(f"usage: {sys.argv[0]} dimM dimN dimK", file=sys.stderr)
+        sys.exit(1)
+    m = int(sys.argv[1])
+    n = int(sys.argv[2])
+    k = int(sys.argv[3])
+    return (m, n, k)
+
+
+# Reorder array in a way that groups two adjacent elements along the column to
+# be now adjacent along the row.  This way, when the resulting fp16 array is
+# read in column-major order with 32-bit granularity, the fp16 elements will be
+# read in the same order as regular fp32 elements in column-major.
+#
+# For example:
+# [[1 2]
+#  [3 4]
+#  [5 6]
+#  [7 8]]
+# becomes
+# [[1 3 2 4]
+#  [5 7 6 8]]
+def pack_fp16_by_column(array):
+    rows = array.shape[0]
+    cols = array.shape[1]
+
+    T = array.transpose([1, 0])
+    T_packed = T.reshape([cols, -1, 2])
+    result = T_packed.transpose([1, 0, 2])
+    return result
+
+
+# Do the same as pack_fp16_by_column, but for every two elements along the row.
+def pack_fp16_by_row(array):
+    rows = array.shape[0]
+    cols = array.shape[1]
+
+    result = array.reshape([rows, -1, 2])
+    return result
+
+
+if __name__ == "__main__":
+    seqlen, _, headdim = parse_mnk()
+
+    rand = True
+    if not rand:
+        A_array = np.arange(seqlen * headdim).reshape([seqlen, headdim])
+        B_array = np.arange(headdim * seqlen).reshape([headdim, seqlen])
+        C_array = np.arange(seqlen * seqlen).reshape([seqlen, headdim])
+    else:
+        np.random.seed(0)
+        A_array = np.random.rand(seqlen, headdim) - 0.5
+        B_array = np.random.rand(headdim, seqlen) - 0.5
+        C_array = np.random.rand(seqlen, headdim) - 0.5
+        # C_array = np.zeros([M, N])
+
+    fp16 = False
+    if fp16:
+        A_packed = pack_fp16_by_row(A_array)
+        AT_packed = A_packed.transpose([1, 0, 2])
+        AT_array = AT_packed.reshape([-1, seqlen * 2])
+        AT_array.astype('float16').tofile("input.a.col.bin")
+        # print('AT:')
+        # print(AT_array)
+        B_packed = pack_fp16_by_column(B_array)
+        B_array = B_packed.reshape([-1, headdim * 2])
+        B_array.astype('float16').tofile("input.b.row.bin")
+        # print('B:')
+        # print(B_array)
+    else:
+        A_array.astype('float32').tofile("input.a.row.bin")
+        AT_array = A_array.transpose([1, 0])
+        AT_array.astype('float32').tofile("input.a.col.bin")
+        B_array.astype('float32').tofile("input.b.bin")
+        C_array.astype('float32').tofile("input.c.bin")
+        # print('AT:')
+        # print(AT_array)
+        # print('B:')
+        # print(B_array)
+
+    assert((seqlen % 64) == 0)
+
+    Br = 64
+    Bc = Br
+
+    rowmax = np.zeros([Br])
+    rowsum = np.zeros([Br])
+    O = np.zeros([Br, headdim])
+
+    def exp2(x):
+        return (x**2) / 2.0 + x + 1.0
+
+    full_S = A_array @ B_array
+    full_S_T = full_S.transpose([1, 0])
+    full_S.astype('float32').tofile("full_S.bin")
+
+    col_to_save = 0
+
+    for col in range(0, seqlen, Bc):
+        print(f"tile iteration {col}~{col + Bc} ======================================")
+
+        # FIXME: only work with the first 64 rows of Q for now
+        Q_tile = A_array[0:64, :]
+        K_tile = B_array[:, col:col+Bc]
+
+        S = Q_tile @ K_tile
+        if col == col_to_save:
+            print('S_expected:')
+            print(S)
+            S.astype('float32').tofile("S_expected.bin")
+
+        # generate rowmax result in online softmax
+        rowmax_this = np.max(S, axis=1)
+        rowmax_prev = rowmax.copy()
+        rowmax = np.maximum(rowmax, rowmax_this)
+        if col == col_to_save:
+            rowmax.astype('float32').tofile("rowmax.bin")
+
+        # subtrace rowmax from each row by broadcasting
+        # (placeholder for exp)
+        x = S - rowmax[:, np.newaxis]
+        P = exp2(x)
+        # for i in range(3, 4):
+        #     P += (x**i) / np.math.factorial(i)
+        # P = np.exp(exp)
+        # print('P error:')
+        # print(P / np.exp(x))
+        if col == col_to_save:
+            print('P_expected:')
+            print(P)
+            P.astype('float32').tofile("P_expected.bin")
+            P.transpose([1, 0]).astype('float32').tofile("P_expected.col.bin")
+
+        rowsum_this = np.sum(P, axis=1)
+        x = rowmax_prev - rowmax_this
+        rowsum = exp2(x) * rowsum + rowsum_this
+        if col == col_to_save:
+            rowsum.astype('float32').tofile("rowsum.bin")
+
+        x = rowmax_prev - rowmax
+        O = O / (exp2(x)[:, np.newaxis])
+        if col == col_to_save:
+            print('O_before_PV:')
+            print(O)
+            O.astype('float32').tofile("O_before_PV.bin")
+
+        V = C_array[col:col+Bc, :]
+        if col == col_to_save:
+            V.astype('float32').tofile("V_expected.bin")
+        # O = P.transpose([1, 0]) @ V
+        O = O + P @ V
+        if col == col_to_save:
+            print('O_after_PV:')
+            print(O)
+            O.astype('float32').tofile("O_after_PV.bin")