tiny.simd.dotF32BFloat16Assume

1pub fn computeF32BFloatAssume(2    comptime D: type,3    a: []const f32,4    b: []const bfloat.BFloat16,5    count: usize,6    comptime assumptions: Assumptions,7) f32 {8    if (comptime D.Lane != f32) @compileError("mixed dot requires an f32 descriptor");9    validateInputs(D.lane_count, a.len, b.len, count, assumptions);10    var sum0: D.Vector = @splat(0);11    var sum1: D.Vector = @splat(0);12    var sum2: D.Vector = @splat(0);13    var sum3: D.Vector = @splat(0);14    var index: usize = 0;15    while (index + 4 * D.lane_count <= count) : (index += 4 * D.lane_count) {16        sum0 = mixedMulAdd(D, a[index..], b[index..], sum0);17        const index1 = index + D.lane_count;18        sum1 = mixedMulAdd(D, a[index1..], b[index1..], sum1);19        const index2 = index + 2 * D.lane_count;20        sum2 = mixedMulAdd(D, a[index2..], b[index2..], sum2);21        const index3 = index + 3 * D.lane_count;22        sum3 = mixedMulAdd(D, a[index3..], b[index3..], sum3);23    }24    while (index + D.lane_count <= count) : (index += D.lane_count) {25        sum0 = mixedMulAdd(D, a[index..], b[index..], sum0);26    }27    if (!assumptions.multiple_of_vector and index != count) {28        const remaining = count - index;29        const av = loadTail(D, a[index..], remaining, assumptions.padded_to_vector);30        const DB = bfloat.Tag(D.lane_count);31        const bits = loadBFloatTail(DB, b[index..], remaining, assumptions.padded_to_vector);32        sum1 = arithmetic.mulAdd(D, av, bfloat.promoteF32(D, bits), sum1);33    }34    return reduce.sum(D, (sum0 + sum1) + (sum2 + sum3));35}

1pub const dotF32BFloat16Assume = dot.computeF32BFloatAssume;